Add Batch 3f11b350-53b1-4c49-bb61-d68b7f5238d9 data

.gitattributes

@@ -6105,3 +6105,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2023/LMGQS_[[:space:]]A[[:space:]]Large-scale[[:space:]]Dataset[[:space:]]for[[:space:]]Query-focused[[:space:]]Summarization/cccbb47a-4240-4c7f-8387-5f19dcfaf91d_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/Language[[:space:]]Guided[[:space:]]Visual[[:space:]]Question[[:space:]]Answering_[[:space:]]Elevate[[:space:]]Your[[:space:]]Multimodal[[:space:]]Language[[:space:]]Model[[:space:]]Using[[:space:]]Knowledge-Enriched[[:space:]]Prompts/b65fb439-6cfa-4e17-ad97-64a399c0e380_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/Language-Agnostic[[:space:]]Bias[[:space:]]Detection[[:space:]]in[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Bias[[:space:]]Probing/e234fdd7-671e-462c-a942-854909126bb1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Large[[:space:]]Language[[:space:]]Model[[:space:]]Is[[:space:]]Not[[:space:]]a[[:space:]]Good[[:space:]]Few-shot[[:space:]]Information[[:space:]]Extractor,[[:space:]]but[[:space:]]a[[:space:]]Good[[:space:]]Reranker[[:space:]]for[[:space:]]Hard[[:space:]]Samples!/8e7fc151-0a2c-4d19-8ed1-5431397bc961_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Large[[:space:]]Language[[:space:]]Models[[:space:]]Are[[:space:]]Better[[:space:]]Adversaries_[[:space:]]Exploring[[:space:]]Generative[[:space:]]Clean-Label[[:space:]]Backdoor[[:space:]]Attacks[[:space:]]Against[[:space:]]Text[[:space:]]Classifiers/0d954e5a-e245-4d04-a6c4-a1f2e1815bb6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Large[[:space:]]Language[[:space:]]Models[[:space:]]Know[[:space:]]Your[[:space:]]Contextual[[:space:]]Search[[:space:]]Intent_[[:space:]]A[[:space:]]Prompting[[:space:]]Framework[[:space:]]for[[:space:]]Conversational[[:space:]]Search/306f5a35-c68f-4d39-a79b-49d0690590d2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Large[[:space:]]Language[[:space:]]Models[[:space:]]Meet[[:space:]]Harry[[:space:]]Potter_[[:space:]]A[[:space:]]Dataset[[:space:]]for[[:space:]]Aligning[[:space:]]Dialogue[[:space:]]Agents[[:space:]]with[[:space:]]Characters/e87156c1-b2b9-4f05-802a-2e45b7de9ff2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Large[[:space:]]Language[[:space:]]Models[[:space:]]are[[:space:]]Better[[:space:]]Reasoners[[:space:]]with[[:space:]]Self-Verification/f1cf3b27-5099-4a74-9aac-c17824766bf4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Large[[:space:]]Language[[:space:]]Models[[:space:]]are[[:space:]]Not[[:space:]]Yet[[:space:]]Human-Level[[:space:]]Evaluators[[:space:]]for[[:space:]]Abstractive[[:space:]]Summarization/4ca6d205-7dab-4c6c-a318-c1a40d34c0a8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Large[[:space:]]Language[[:space:]]Models[[:space:]]as[[:space:]]Source[[:space:]]Planner[[:space:]]for[[:space:]]Personalized[[:space:]]Knowledge-grounded[[:space:]]Dialogues/3db068d7-7569-4325-831f-81a84ff080cc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Large-Scale[[:space:]]and[[:space:]]Multi-Perspective[[:space:]]Opinion[[:space:]]Summarization[[:space:]]with[[:space:]]Diverse[[:space:]]Review[[:space:]]Subsets/4b542fcd-babd-4d16-9e9d-894d20925ab6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Late[[:space:]]Fusion[[:space:]]of[[:space:]]Transformers[[:space:]]for[[:space:]]Sentiment[[:space:]]Analysis[[:space:]]of[[:space:]]Code-Switched[[:space:]]Data/61718999-dcc2-4475-b2ab-d1733664f853_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/LayoutDIT_[[:space:]]Layout-Aware[[:space:]]End-to-End[[:space:]]Document[[:space:]]Image[[:space:]]Translation[[:space:]]with[[:space:]]Multi-Step[[:space:]]Conductive[[:space:]]Decoder/cc5f0dc7-8232-4817-8945-d5c4f4e97dcf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learn[[:space:]]From[[:space:]]One[[:space:]]Specialized[[:space:]]Sub-Teacher_[[:space:]]One-to-One[[:space:]]Mapping[[:space:]]for[[:space:]]Feature-Based[[:space:]]Knowledge[[:space:]]Distillation/60cdbc81-d656-43bb-933a-bfbb84e0d79b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learn[[:space:]]Your[[:space:]]Tokens_[[:space:]]Word-Pooled[[:space:]]Tokenization[[:space:]]for[[:space:]]Language[[:space:]]Modeling/0366ee09-2bfc-4dc2-9325-6bd4cad5bcfb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Dynamic[[:space:]]Representations[[:space:]]for[[:space:]]Discourse[[:space:]]Dependency[[:space:]]Parsing/abc4a4db-8642-416a-bfe3-ed2988f2b01d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Easily[[:space:]]Updated[[:space:]]General[[:space:]]Purpose[[:space:]]Text[[:space:]]Representations[[:space:]]with[[:space:]]Adaptable[[:space:]]Task-Specific[[:space:]]Prefix/286adf1c-ffaf-4daa-a510-239a5e12b99d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Interpretable[[:space:]]Style[[:space:]]Embeddings[[:space:]]via[[:space:]]Prompting[[:space:]]LLMs/33cb939d-a377-48e9-a884-0f46977e1a6d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]Semantic[[:space:]]Role[[:space:]]Labeling[[:space:]]from[[:space:]]Compatible[[:space:]]Label[[:space:]]Sequences/25f4c238-0ca4-4093-a80e-5a78cfd2f7c5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]to[[:space:]]Abstract[[:space:]]with[[:space:]]Nonparametric[[:space:]]Variational[[:space:]]Information[[:space:]]Bottleneck/c41be780-b9f6-486f-ae93-4315ba6c70e6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]to[[:space:]]Compose[[:space:]]Representations[[:space:]]of[[:space:]]Different[[:space:]]Encoder[[:space:]]Layers[[:space:]]towards[[:space:]]Improving[[:space:]]Compositional[[:space:]]Generalization/13a15a00-fbc8-484f-82ef-01ad47b682e3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]to[[:space:]]Correct[[:space:]]Noisy[[:space:]]Labels[[:space:]]for[[:space:]]Fine-Grained[[:space:]]Entity[[:space:]]Typing[[:space:]]via[[:space:]]Co-Prediction[[:space:]]Prompt[[:space:]]Tuning/6bd1b093-eb88-4fbc-9ad7-760ac5227622_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]to[[:space:]]Follow[[:space:]]Object-Centric[[:space:]]Image[[:space:]]Editing[[:space:]]Instructions[[:space:]]Faithfully/1d551fc3-ecbd-4738-9c52-97ec318510df_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]to[[:space:]]love[[:space:]]diligent[[:space:]]trolls_[[:space:]]Accounting[[:space:]]for[[:space:]]rater[[:space:]]effects[[:space:]]in[[:space:]]the[[:space:]]dialogue[[:space:]]safety[[:space:]]task/30afa8a9-1898-4f0d-a0eb-521fee637d6a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Learning[[:space:]]under[[:space:]]Label[[:space:]]Proportions[[:space:]]for[[:space:]]Text[[:space:]]Classification/c7ed2103-e89a-4bbc-82b1-3a49f1b6688f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Legally[[:space:]]Enforceable[[:space:]]Hate[[:space:]]Speech[[:space:]]Detection[[:space:]]for[[:space:]]Public[[:space:]]Forums/eb550035-1c57-43a3-b1fc-960667ef4c64_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Length-Adaptive[[:space:]]Distillation_[[:space:]]Customizing[[:space:]]Small[[:space:]]Language[[:space:]]Model[[:space:]]for[[:space:]]Dynamic[[:space:]]Token[[:space:]]Pruning/f1d0235f-a268-417d-9849-4dd40e514268_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Less[[:space:]]than[[:space:]]One-shot_[[:space:]]Named[[:space:]]Entity[[:space:]]Recognition[[:space:]]via[[:space:]]Extremely[[:space:]]Weak[[:space:]]Supervision/9730ccef-0f86-4695-8a9e-ca4ebf3a5d88_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Let’s[[:space:]]Synthesize[[:space:]]Step[[:space:]]by[[:space:]]Step_[[:space:]]Iterative[[:space:]]Dataset[[:space:]]Synthesis[[:space:]]with[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]by[[:space:]]Extrapolating[[:space:]]Errors[[:space:]]from[[:space:]]Small[[:space:]]Models/5ffb9959-bfd0-489c-93b8-bc176a7556ed_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Leveraging[[:space:]]Contrastive[[:space:]]Learning[[:space:]]and[[:space:]]Knowledge[[:space:]]Distillation[[:space:]]for[[:space:]]Incomplete[[:space:]]Modality[[:space:]]Rumor[[:space:]]Detection/62ef1015-49e4-479c-a34d-d7f4b65661bb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Leveraging[[:space:]]GPT-4[[:space:]]for[[:space:]]Automatic[[:space:]]Translation[[:space:]]Post-Editing/d1da0d75-71ef-4088-8ded-cbcf7886c2b4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Leveraging[[:space:]]Multiple[[:space:]]Teachers[[:space:]]for[[:space:]]Test-Time[[:space:]]Adaptation[[:space:]]of[[:space:]]Language-Guided[[:space:]]Classifiers/2ec6d0c2-6047-4986-a768-d4cdd95a1cb4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Leveraging[[:space:]]Structured[[:space:]]Information[[:space:]]for[[:space:]]Explainable[[:space:]]Multi-hop[[:space:]]Question[[:space:]]Answering[[:space:]]and[[:space:]]Reasoning/4006a8f2-0d5d-4077-b4c8-82164c6f8a49_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Lexical[[:space:]]Entrainment[[:space:]]for[[:space:]]Conversational[[:space:]]Systems/3d2505b6-a61e-4fa1-906b-5714d21ad360_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Lexical[[:space:]]Repetitions[[:space:]]Lead[[:space:]]to[[:space:]]Rote[[:space:]]Learning_[[:space:]]Unveiling[[:space:]]the[[:space:]]Impact[[:space:]]of[[:space:]]Lexical[[:space:]]Overlap[[:space:]]in[[:space:]]Train[[:space:]]and[[:space:]]Test[[:space:]]Reference[[:space:]]Summaries/0a1809ab-3a08-4ede-bff4-8a07c4227603_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Licon_[[:space:]]A[[:space:]]Diverse,[[:space:]]Controllable[[:space:]]and[[:space:]]Challenging[[:space:]]Linguistic[[:space:]]Concept[[:space:]]Learning[[:space:]]Benchmark/02eadb72-d45d-4f13-9501-be9a3b0c6126_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Linguistic[[:space:]]Compression[[:space:]]in[[:space:]]Single-Sentence[[:space:]]Human-Written[[:space:]]Summaries/8fd47e5d-acc0-4e85-9cc1-81c88e35102f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Linguistically[[:space:]]Motivated[[:space:]]Sign[[:space:]]Language[[:space:]]Segmentation/390cab7a-e6c0-4361-a52a-d7714d5d2e03_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Locally[[:space:]]Differentially[[:space:]]Private[[:space:]]Document[[:space:]]Generation[[:space:]]Using[[:space:]]Zero[[:space:]]Shot[[:space:]]Prompting/7a6ec7bc-59a2-4f94-879f-d38371f96872_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/LogiCoT_[[:space:]]Logical[[:space:]]Chain-of-Thought[[:space:]]Instruction[[:space:]]Tuning/1d9b901b-71fa-4d39-b392-2fb761b86ede_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Logic[[:space:]]Unveils[[:space:]]Truth,[[:space:]]While[[:space:]]Disguise[[:space:]]Obscures[[:space:]]It_[[:space:]]Transition[[:space:]]Logic[[:space:]]Augmented[[:space:]]Response[[:space:]]Selection[[:space:]]for[[:space:]]Multi-Turn[[:space:]]Dialogue/69a3b67d-60d2-478c-8145-48f38e25e50c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Logic-LM_[[:space:]]Empowering[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Symbolic[[:space:]]Solvers[[:space:]]for[[:space:]]Faithful[[:space:]]Logical[[:space:]]Reasoning/ab17e310-b23c-4108-acbe-46ce22d6b93c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/LogicAttack_[[:space:]]Adversarial[[:space:]]Attacks[[:space:]]for[[:space:]]Evaluating[[:space:]]Logical[[:space:]]Consistency[[:space:]]of[[:space:]]Natural[[:space:]]Language[[:space:]]Inference/f746a610-5233-4b58-902c-78e21324fd77_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Long-Form[[:space:]]Speech[[:space:]]Translation[[:space:]]through[[:space:]]Segmentation[[:space:]]with[[:space:]]Finite-State[[:space:]]Decoding[[:space:]]Constraints[[:space:]]on[[:space:]]Large[[:space:]]Language[[:space:]]Models/d3e17d00-522f-4070-a45a-0ae350d2ce2b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Long-Horizon[[:space:]]Dialogue[[:space:]]Understanding[[:space:]]for[[:space:]]Role[[:space:]]Identification[[:space:]]in[[:space:]]the[[:space:]]Game[[:space:]]of[[:space:]]Avalon[[:space:]]with[[:space:]]Large[[:space:]]Language[[:space:]]Models/21a867f4-5d35-40f8-8202-4e1c9d7a3ece_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Long-Range[[:space:]]Language[[:space:]]Modeling[[:space:]]with[[:space:]]Selective[[:space:]]Cache/05301bd4-67ef-49fc-87c4-f6d0f147f387_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Loose[[:space:]]lips[[:space:]]sink[[:space:]]ships_[[:space:]]Mitigating[[:space:]]Length[[:space:]]Bias[[:space:]]in[[:space:]]Reinforcement[[:space:]]Learning[[:space:]]from[[:space:]]Human[[:space:]]Feedback/9a4f3731-40f7-4621-ad7d-89956b5194dd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Low-Resource[[:space:]]Comparative[[:space:]]Opinion[[:space:]]Quintuple[[:space:]]Extraction[[:space:]]by[[:space:]]Data[[:space:]]Augmentation[[:space:]]with[[:space:]]Prompting/ae11d62d-597c-4442-84ff-cbba02a8f44d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/M2C_[[:space:]]Towards[[:space:]]Automatic[[:space:]]Multimodal[[:space:]]Manga[[:space:]]Complement/dd7f70bf-284c-4189-a881-de5226139600_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MAPO_[[:space:]]Boosting[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]Performance[[:space:]]with[[:space:]]Model-Adaptive[[:space:]]Prompt[[:space:]]Optimization/576fc35f-9418-4463-a828-bde0e3351871_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MCC-KD_[[:space:]]Multi-CoT[[:space:]]Consistent[[:space:]]Knowledge[[:space:]]Distillation/a59ae8b4-7ad5-4107-91a6-aabacd3921bd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MCLF_[[:space:]]A[[:space:]]Multi-grained[[:space:]]Contrastive[[:space:]]Learning[[:space:]]Framework[[:space:]]for[[:space:]]ASR-robust[[:space:]]Spoken[[:space:]]Language[[:space:]]Understanding/218e43d6-faeb-4fbd-8e99-40d724b368d1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MEAL_[[:space:]]Stable[[:space:]]and[[:space:]]Active[[:space:]]Learning[[:space:]]for[[:space:]]Few-Shot[[:space:]]Prompting/b8821ec0-a306-4d8d-9ec2-fda925d19126_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MEEP_[[:space:]]Is[[:space:]]this[[:space:]]Engaging_[[:space:]]Prompting[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Dialogue[[:space:]]Evaluation[[:space:]]in[[:space:]]Multilingual[[:space:]]Settings/3fbb3463-ceb8-4dc2-8df8-75a38d871198_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MEGClass_[[:space:]]Extremely[[:space:]]Weakly[[:space:]]Supervised[[:space:]]Text[[:space:]]Classification[[:space:]]via[[:space:]]Mutually-Enhancing[[:space:]]Text[[:space:]]Granularities/6407abb6-f372-42f9-9ecb-648c988e4f29_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MISCA_[[:space:]]A[[:space:]]Joint[[:space:]]Model[[:space:]]for[[:space:]]Multiple[[:space:]]Intent[[:space:]]Detection[[:space:]]and[[:space:]]Slot[[:space:]]Filling[[:space:]]with[[:space:]]Intent-Slot[[:space:]]Co-Attention/5f71055d-de18-42cb-998d-ec4de7789c48_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MM-Reasoner_[[:space:]]A[[:space:]]Multi-Modal[[:space:]]Knowledge-Aware[[:space:]]Framework[[:space:]]for[[:space:]]Knowledge-Based[[:space:]]Visual[[:space:]]Question[[:space:]]Answering/95e53a92-859e-400c-a84f-604e9b8a7fc2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MPrompt_[[:space:]]Exploring[[:space:]]Multi-level[[:space:]]Prompt[[:space:]]Tuning[[:space:]]for[[:space:]]Machine[[:space:]]Reading[[:space:]]Comprehension/b032ad9c-ebc4-4ecb-863d-d00748cbbfc5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MRRL_[[:space:]]Modifying[[:space:]]the[[:space:]]Reference[[:space:]]via[[:space:]]Reinforcement[[:space:]]Learning[[:space:]]for[[:space:]]Non-Autoregressive[[:space:]]Joint[[:space:]]Multiple[[:space:]]Intent[[:space:]]Detection[[:space:]]and[[:space:]]Slot[[:space:]]Filling/00146739-cadc-4952-995c-b28125446516_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MSCFFN_[[:space:]]A[[:space:]]New[[:space:]]FFN[[:space:]]with[[:space:]]Multi-Space[[:space:]]Cross[[:space:]]to[[:space:]]Accelerate[[:space:]]Transformer/98328ffe-b339-49a9-b1b2-f66faae994ba_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MTGER_[[:space:]]Multi-view[[:space:]]Temporal[[:space:]]Graph[[:space:]]Enhanced[[:space:]]Temporal[[:space:]]Reasoning[[:space:]]over[[:space:]]Time-Involved[[:space:]]Document/40b776c4-358a-4c45-a2bc-569db9a17697_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MUX-PLMs_[[:space:]]Data[[:space:]]Multiplexing[[:space:]]for[[:space:]]High-throughput[[:space:]]Language[[:space:]]Models/a27f6406-030a-435a-812b-a48eaa129e9c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MWE[[:space:]]as[[:space:]]WSD_[[:space:]]Solving[[:space:]]Multiword[[:space:]]Expression[[:space:]]Identification[[:space:]]with[[:space:]]Word[[:space:]]Sense[[:space:]]Disambiguation/594deecf-a580-43e7-b5d8-3c309dc0f756_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MaXM_[[:space:]]Towards[[:space:]]Multilingual[[:space:]]Visual[[:space:]]Question[[:space:]]Answering/bf3fe977-b566-451c-a6c4-1f46b280df9f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MacLaSa_[[:space:]]Multi-Aspect[[:space:]]Controllable[[:space:]]Text[[:space:]]Generation[[:space:]]via[[:space:]]Efficient[[:space:]]Sampling[[:space:]]from[[:space:]]Compact[[:space:]]Latent[[:space:]]Space/45a8f5b4-d87f-434e-a5ac-c0b03234c505_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Macedon_[[:space:]]Minimizing[[:space:]]Representation[[:space:]]Coding[[:space:]]Rate[[:space:]]Reduction[[:space:]]for[[:space:]]Cross-Lingual[[:space:]]Natural[[:space:]]Language[[:space:]]Understanding/e15e7dc3-c861-4c4f-aea2-b06f9e5aa84a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Machine[[:space:]]Reading[[:space:]]Comprehension[[:space:]]using[[:space:]]Case-based[[:space:]]Reasoning/180b5324-8a41-4f6e-98a2-d090f2a5d8be_origin.pdf filter=lfs diff=lfs merge=lfs -text

2023/Large Language Model Is Not a Good Few-shot Information Extractor, but a Good Reranker for Hard Samples!/8e7fc151-0a2c-4d19-8ed1-5431397bc961_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ef40298a5041cb6e5dc211e4a604b4120a889b2abebfd0e2860e85af231673f
+size 1440454

2023/Large Language Model Is Not a Good Few-shot Information Extractor, but a Good Reranker for Hard Samples!/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5f60fa8729de43c9d5c8cd159d217cb7a9f020e01de790f649a06cbbf48bb390
+size 3620958

2023/Large Language Models Are Better Adversaries_ Exploring Generative Clean-Label Backdoor Attacks Against Text Classifiers/0d954e5a-e245-4d04-a6c4-a1f2e1815bb6_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a7b968da7e7c1bb67f3fbd58eaa86f8264703da86e7f645e888847dc5520efc9
+size 985875

2023/Large Language Models Are Better Adversaries_ Exploring Generative Clean-Label Backdoor Attacks Against Text Classifiers/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c6dfd94d3de49bbb3b129ef88b57b502df84bbec78439a6622c87043df9dd6a
+size 2652009

2023/Large Language Models Know Your Contextual Search Intent_ A Prompting Framework for Conversational Search/306f5a35-c68f-4d39-a79b-49d0690590d2_content_list.json

@@ -0,0 +1,2288 @@
+[
+    {
+        "type": "text",
+        "text": "Large Language Models Know Your Contextual Search Intent: A Prompting Framework for Conversational Search",
+        "text_level": 1,
+        "bbox": [
+            174,
+            79,
+            825,
+            118
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Kelong Mao $^{1,2}$ , Zhicheng Dou $^{1,2*}$ , Fengran Mo $^{3}$ , Jiewen Hou $^{4}$ , Haonan Chen $^{1}$ , Hongjin Qian $^{1}$",
+        "bbox": [
+            236,
+            124,
+            763,
+            158
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>1</sup>Gaoling School of Artificial Intelligence, Renmin University of China",
+        "bbox": [
+            213,
+            159,
+            786,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{2}$ Engineering Research Center of Next-Generation Search and Recommendation, MOE",
+        "bbox": [
+            149,
+            175,
+            852,
+            191
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>3</sup>Université de Montréal, Québec, Canada",
+        "bbox": [
+            331,
+            192,
+            670,
+            206
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{4}$ Institute of Computing Technology, Chinese Academy of Sciences",
+        "bbox": [
+            228,
+            209,
+            773,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{mk1,dou}@ruc.edu.cn",
+        "bbox": [
+            400,
+            225,
+            601,
+            241
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            266
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Precisely understanding users' contextual search intent has been an important challenge for conversational search. As conversational search sessions are much more diverse and long-tailed, existing methods trained on limited data still show unsatisfactory effectiveness and robustness to handle real conversational search scenarios. Recently, large language models (LLMs) have demonstrated amazing capabilities for text generation and conversation understanding. In this work, we present a simple yet effective prompting framework, called LLM4CS, to leverage LLMs as a text-based search intent interpreter to help conversational search. Under this framework, we explore three prompting methods to generate multiple query rewrites and hypothetical responses, and propose to aggregate them into an integrated representation that can robustly represent the user's real contextual search intent. Extensive automatic evaluations and human evaluations on three widely used conversational search benchmarks, including CAsT-19, CAsT-20, and CAsT-21, demonstrate the remarkable performance of our simple LLM4CS framework compared with existing methods and even using human rewrites. Our findings provide important evidence to better understand and leverage LLMs for conversational search. The code is released at https://github.com/kyriemao/LLM4CS.",
+        "bbox": [
+            144,
+            278,
+            460,
+            717
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            728,
+            258,
+            743
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Conversational search has been expected to be the next generation of search paradigms (Culpepper et al., 2018). It supports search via conversation to provide users with more accurate and intuitive search results and a much more user-friendly search experience. Unlike using traditional search engines which mainly process keyword queries, users could imagine the conversational search system as a knowledgeable human expert and directly start a",
+        "bbox": [
+            112,
+            753,
+            487,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "multi-turn conversation with it in natural languages to solve their questions. However, one of the main challenges for this beautiful vision is that the users' queries may contain some linguistic problems (e.g., omissions and coreference) and it becomes much harder to capture their real search intent under the multi-turn conversation context (Dalton et al., 2021; Mao et al., 2022a).",
+        "bbox": [
+            507,
+            253,
+            884,
+            381
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To achieve conversational search, an intuitive method known as Conversational Query Rewriting (CQR) involves using a rewriting model to transform the current query into a de-contextualized form. Subsequently, any ad-hoc search models can be seamlessly applied for retrieval purposes. Given that existing ad-hoc search models can be reused directly, CQR demonstrates substantial practical value for industries in quickly initializing their conversational search engines. Another type of method, Conversational Dense Retrieval (CDR), tries to learn a conversational dense retriever to encode the user's real search intent and passages into latent representations and performs dense retrieval. In contrast to the two-step CQR method, where the rewriter is difficult to be directly optimized towards search (Yu et al., 2021; Mao et al., 2023a), the conversational dense retriever can naturally learn from session-passage relevance signals.",
+        "bbox": [
+            507,
+            384,
+            884,
+            690
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "However, as conversational search sessions are much more diverse and long-tailed (Mao et al., 2022b; Dai et al., 2022; Mo et al., 2023a), existing CQR and CDR methods trained on limited data still show unsatisfactory performance, especially on more complex conversational search sessions. Many studies (Vakulenko et al., 2021b; Lin et al., 2021a; Qian and Dou, 2022; Krasakis et al., 2022) have demonstrated the performance advantages of using de-contextualized human rewrites on sessions which have complex response dependency. Also, as reported in the public TREC CAst 2021 benchmark (Dalton et al., 2022), existing methods still suffer from significant degradation in their ef",
+        "bbox": [
+            507,
+            694,
+            884,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding author.",
+        "bbox": [
+            141,
+            904,
+            287,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "1211",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 1211-1225 December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            216,
+            945,
+            779,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "fectiveness as conversations become longer.",
+        "bbox": [
+            114,
+            85,
+            442,
+            99
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Recently, large language models (LLMs) have shown amazing capabilities for text generation and conversation understanding (Brown et al., 2020; Wei et al., 2022; Thoppilan et al., 2022; Zhu et al., 2023). In the field of information retrieval (IR), LLMs have also been successfully utilized to enhance relevance modeling via various techniques such as query generation (Bonifacio et al., 2022; Dai et al., 2023), query expansion (Wang et al., 2023a), document prediction (Gao et al., 2022; Mackie et al., 2023), etc. Inspired by the strong performance of LLMs in conversation and IR, we try to investigate how LLMs can be adapted to precisely grasp users' contextual search intent for conversational search.",
+        "bbox": [
+            112,
+            101,
+            489,
+            340
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this work, we present a simple yet effective prompting framework, called LLM4CS, to leverage LLM as a search intent interpreter to facilitate conversational search. Specifically, we first prompt LLM to generate both short query rewrites and longer hypothetical responses in multiple perspectives and then aggregate these generated contents into an integrated representation that robustly represents the user's real search intent. Under our framework, we propose three specific prompting methods and aggregation methods, and conduct extensive evaluations on three widely used conversational search benchmarks, including CAsT-19 (Dalton et al., 2020), CAsT-20 (Dalton et al., 2021), and CAsT-21 (Dalton et al., 2022), to comprehensively investigate the effectiveness of LLMs for conversational search.",
+        "bbox": [
+            112,
+            343,
+            489,
+            615
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In general, our framework has two main advantages. First, by leveraging the powerful contextual understanding and generation abilities of large language models, we show that additionally generating hypothetical responses to explicitly supplement more plausible search intents underlying the short rewrite can significantly improve the search performance. Second, we show that properly aggregating multiple rewrites and hypothetical responses can effectively filter out incorrect search intents and enhance the reasonable ones, leading to better search performance and robustness.",
+        "bbox": [
+            112,
+            617,
+            489,
+            809
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Overall, our main contributions are:",
+        "bbox": [
+            132,
+            810,
+            401,
+            826
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "- We propose a prompting framework and design three tailored prompting methods to leverage large language models for conversational search, which effectively circumvents the serious data scarcity problem faced by the con",
+        "bbox": [
+            134,
+            839,
+            489,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/f990b9bc6f9a08ce86f9fc2c76a6d598b24e489fcc8578691316fe54a36a4f90.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            608,
+            82,
+            647,
+            107
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Context & Question",
+        "bbox": [
+            650,
+            87,
+            779,
+            101
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/8548264f52201c30c17392b3663df4e0506a6f0c858cc5d28141700383dc3db9.jpg",
+        "image_caption": [
+            "Figure 1: An overview of LLM4CS."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            557,
+            118,
+            863,
+            414
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "versational search field.",
+        "bbox": [
+            544,
+            456,
+            724,
+            469
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We show that additionally generating hypothetical responses and properly aggregating multiple generated results are crucial for improving search performance.",
+            "- We demonstrate the exceptional effectiveness of LLMs for conversational search through both automatic and human evaluations, where the best method in our LLM4CS achieves remarkable improvements in search performance over state-of-the-art CQR and CDR baselines, surpassing even human rewrites."
+        ],
+        "bbox": [
+            531,
+            483,
+            884,
+            673
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            507,
+            699,
+            665,
+            714
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Conversational Search. Conversational search is an evolving field that involves retrieving relevant information based on multi-turn dialogues with users. To achieve conversational search, two main methods have been developed: conversational query rewriting and conversational dense retrieval. Conversational query rewriting converts the conversational search problem into an ad-hoc search problem by reformulating the search session into a standalone query rewrite. Existing methods try to select useful tokens from the conversation context (Voskarides et al., 2020; Lin et al., 2021b)",
+        "bbox": [
+            507,
+            726,
+            884,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "1212",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "or train a generative rewritten based on the pairs of sessions and rewrites (Lin et al., 2020; Yu et al., 2020; Vakulenko et al., 2021a). To make the rewriting process aware of the downstream retrieval process, some studies propose to adopt reinforcement learning (Wu et al., 2022; Chen et al., 2022) or enhance the learning of rewritten with ranking signals (Mao et al., 2023a; Mo et al., 2023a). On the other hand, conversational dense retrieval (Yu et al., 2021) directly encodes the whole conversational search session to perform end-to-end dense retrieval. Existing methods mainly try to improve the session representation through context denoising (Mao et al., 2022a; Krasakis et al., 2022; Mo et al., 2023b; Mao et al., 2023b), data augmentation (Lin et al., 2021a; Mao et al., 2022b; Dai et al., 2022), and hard negative mining (Kim and Kim, 2022).",
+        "bbox": [
+            110,
+            84,
+            492,
+            375
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "IR with LLMs. Due to the revolutionary natural language understanding and generation abilities, LLMs are attracting more and more attention from the IR community. LLMs have been leveraged to enhance the relevance modeling of retrieval through query generation (Bonifacio et al., 2022; Jeronymo et al., 2023; Dai et al., 2023), query expansion (Wang et al., 2023a), document prediction (Gao et al., 2022; Mackie et al., 2023), etc. Besides, Shen et al. (2023) proposed to first use the retriever to enhance the generation of LLM and then use the generated content to augment the original search query for better retrieval. Ziems et al. (2023) treated LLM as a built-in search engine to retrieve documents based on the generated URL. There are also some works leveraging LLM to perform re-ranking (Sun et al., 2023; Jiang et al., 2023). Different from previous studies, in this paper, we propose the LLM4CS framework that focuses on studying how LLM can be well utilized to capture the user's contextual search intent to facilitate conversational search.",
+        "bbox": [
+            110,
+            391,
+            490,
+            744
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 LLM4CS: Prompting Large Language Models for Conversational Search",
+        "text_level": 1,
+        "bbox": [
+            112,
+            760,
+            482,
+            793
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we introduce our LLM4CS framework, which leverages LLM as a text-based search intent interpreter to facilitate conversational search. Figure 1 shows an overview of LLM4CS. In the following, we first describe our task formulation of conversational search, and then we elaborate on the specific prompting methods and aggregation",
+        "bbox": [
+            112,
+            806,
+            490,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "methods integrated into the framework. Finally, we introduce the retrieval process.",
+        "bbox": [
+            507,
+            84,
+            882,
+            116
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Task Formulation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            128,
+            697,
+            142
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We focus on the task of conversational passage retrieval, which is the crucial first step of conversational search that helps the model access the right evidence knowledge. Given the user query  $q^t$  and the conversation context  $C^t = (q^1, r^1, \\dots, q^{t-1}, r^{t-1})$  of the current turn  $t$ , where  $q^i$  and  $r^i$  denote the user query and the system response of the historical  $i$ -th turn, our goal is to retrieve passages that are relevant to satisfy the user's real search intent of the current turn.",
+        "bbox": [
+            507,
+            149,
+            884,
+            309
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Prompting Methods",
+        "text_level": 1,
+        "bbox": [
+            507,
+            323,
+            714,
+            338
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The prompt follows the formulation of [Instruction, Demonstrations, Input], where Input is composed of the query  $q^t$  and the conversation context  $C^t$  of the current turn  $t$ . Figure 4 shows a general illustration of the prompt construction. Specifically, we design and explore three prompting methods, including Rewriting (REW), Rewriting-Then-Response (RTR), and Rewriting-And-Response (RAR), in our LLM4CS framework.",
+        "bbox": [
+            507,
+            344,
+            884,
+            489
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2.1 Rewriting Prompt (REW)",
+        "text_level": 1,
+        "bbox": [
+            507,
+            499,
+            774,
+            514
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this prompting method, we directly treat LLM as a well-trained conversational query rewriter and prompt it to generate rewrites. Only the red part of Figure 4 is enabled. Although straightforward, we show in Section 4.5 that this simple prompting method has been able to achieve quite a strong search performance compared to existing baselines.",
+        "bbox": [
+            507,
+            519,
+            882,
+            631
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2.2 Rewriting-Then-Response (RTR)",
+        "text_level": 1,
+        "bbox": [
+            507,
+            642,
+            830,
+            657
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Recently, a few studies (Mao et al., 2021; Gao et al., 2022; Yu et al., 2023; Mackie et al., 2023) have shown that generating hypothetical responses for search queries can often bring positive improvements in retrieval performance. Inspired by them, in addition to prompting LLM to generate rewrites, we continue to utilize the generated rewrites to further prompt LLM to generate hypothetical responses that may contain relevant information to answer the current question. The orange part and the blue part of Figure 4 are enabled. Specifically, we incorporate the pre-generated rewrite (i.e., the orange part) into the Input field of the prompt and",
+        "bbox": [
+            507,
+            662,
+            882,
+            871
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>We put this figure in Appendix A due to the space limitation. See our open-sourced code for the full prompt of each prompting method.",
+        "bbox": [
+            507,
+            881,
+            882,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "1213",
+        "bbox": [
+            482,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "then prompt LLM to generate informative hypothetical responses by referring to the rewrite.",
+        "bbox": [
+            112,
+            84,
+            489,
+            116
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2.3 Rewriting-And-Response (RAR)",
+        "text_level": 1,
+        "bbox": [
+            112,
+            124,
+            428,
+            140
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Instead of generating rewrites and hypothetical responses in a two-stage manner, we can also generate them all at once with the red part and the blue part of Figure 4 being enabled. We try to explore whether this one-stage generation could lead to better consistency and accuracy between the generated rewrites and responses, compared with the two-step RTR method.",
+        "bbox": [
+            112,
+            143,
+            487,
+            271
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2.4 Incorporating Chain-of-Thought",
+        "text_level": 1,
+        "bbox": [
+            112,
+            280,
+            431,
+            296
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Chain-of-thought (CoT) (Wei et al., 2020) induces the large language models to decompose a reasoning task into multiple intermediate steps which can unlock their stronger reasoning abilities. In this work, we also investigate whether incorporating the chain-of-thought of reasoning the user's real search intent could improve the quality of rewrite and response generation.",
+        "bbox": [
+            112,
+            299,
+            487,
+            426
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Specifically, as shown in the green part of Figure 4, we manually write the chain-of-thought for each turn of the demonstration, which reflects how humans infer the user's real search intent of the current turn based on the historical conversation context. When generating, we instruct LLM to first generate the chain-of-thought before generating rewrites (and responses). We investigate the effects of our proposed CoT tailored to the reasoning of contextual search intent in Section 4.6.",
+        "bbox": [
+            112,
+            429,
+            489,
+            588
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3 Content Aggregation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            599,
+            326,
+            615
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "After prompting LLM multiple times to generate multiple rewrites and hypothetical responses, we then aggregate these generated contents into an integrated representation to represent the user's complete search intent for search. Let us consider that we have generated  $N$  query rewrites  $\\mathcal{Q} = (\\hat{q}_1,\\dots,\\hat{q}_N)$  and  $M$  hypothetical responses  $\\mathcal{R} = (\\hat{r}_{i1},\\dots,\\hat{r}_{iM})$  for each rewrite  $\\hat{q}_i$ , sorted by their generation probabilities from high to low<sup>2</sup>. Note that in RAR prompting, the rewrites and the hypothetical responses are always generated in pairs (i.e.,  $M = 1$ ). While in RTR prompting, one rewrite can have  $M$  hypothetical responses since they are generated in a two-stage manner. Next, we utilize a dual well-trained ad-hoc retriever<sup>3</sup>  $f$",
+        "bbox": [
+            112,
+            619,
+            489,
+            860
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "(e.g, ANCE (Xiong et al., 2021)) to encode each of them into a high-dimensional intent vector and aggregate these intent vectors into one final search intent vector s. Specifically, we design and explore the following three aggregation methods, including MaxProb, Self-Consistency (SC), and Mean, in our LLM4CS framework.",
+        "bbox": [
+            507,
+            84,
+            882,
+            197
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3.1 MaxProb",
+        "text_level": 1,
+        "bbox": [
+            507,
+            209,
+            645,
+            224
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We directly use the rewrite and the hypothetical response that have the highest generation probabilities. Therefore, compared with the other two aggregation methods that will be introduced later, MaxProb is highly efficient since it actually does not require multiple generations.",
+        "bbox": [
+            507,
+            231,
+            882,
+            326
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Formally, for REW prompting:",
+        "bbox": [
+            527,
+            329,
+            761,
+            344
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {s} = f \\left(\\hat {q} _ {1}\\right). \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            635,
+            361,
+            882,
+            378
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For the RTR and RAR prompting methods, we mix the rewrite and hypothetical response vectors:",
+        "bbox": [
+            507,
+            395,
+            882,
+            426
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {s} = \\frac {f (\\hat {q} _ {1}) + f (\\hat {r} _ {1 1})}{2}. \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            598,
+            441,
+            882,
+            473
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3.2 Self-Consistency (SC)",
+        "text_level": 1,
+        "bbox": [
+            507,
+            488,
+            739,
+            504
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The multiple generated rewrites and hypothetical responses may express different search intents but only some of them are correct. To obtain a more reasonable and consistent search intent representation, we extend the self-consistency prompting method (Wang et al., 2023b), which was initially designed for reasoning tasks with predetermined answer sets, to our contextual search intent understanding task, which lacks a fixed standard answer. To be specific, we select the intent vector that is the most similar to the cluster center of all intent vectors as the final search intent vector, since it represents the most popular search intent overall.",
+        "bbox": [
+            507,
+            508,
+            882,
+            718
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Formally, for REW prompting:",
+        "bbox": [
+            526,
+            719,
+            761,
+            734
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {\\mathbf {q}} ^ {*} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} f (\\hat {q} _ {i}), \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            579,
+            750,
+            882,
+            793
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {s} = \\arg \\max  _ {f (\\hat {q} _ {i})} f (\\hat {q} _ {i}) ^ {\\top} \\cdot \\hat {\\mathbf {q}} ^ {*}, \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            591,
+            795,
+            882,
+            822
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\hat{\\mathbf{q}}^*$  is the cluster center vector and  $\\cdot$  denotes the dot product that measures the similarity.",
+        "bbox": [
+            507,
+            838,
+            880,
+            868
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For RTR prompting, we first select the intent vector  $f(\\hat{q}_k)$  and then select the intent vector  $f(\\hat{r}_{kz})$  from all hypothetical responses generated based on",
+        "bbox": [
+            507,
+            871,
+            882,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "2That is, the generation probability orders are:  $P(\\hat{q}_1)\\geq$ $\\dots \\geq P(\\hat{q}_N)$  and  $P(\\hat{r}_{i1})\\geq \\ldots \\geq P(\\hat{r}_{iM})$",
+        "bbox": [
+            112,
+            866,
+            487,
+            894
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "3The parameters of the query encoder and the passage encoder are shared.",
+        "bbox": [
+            112,
+            892,
+            485,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "1214",
+        "bbox": [
+            482,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "the selected rewrite  $\\hat{q}_k$ :",
+        "bbox": [
+            112,
+            85,
+            287,
+            99
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nk = \\arg \\max _ {i} f (\\hat {q} _ {i}) ^ {\\top} \\cdot \\hat {\\mathbf {q}} ^ {*}, \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            189,
+            110,
+            487,
+            136
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {\\mathbf {r}} _ {k} ^ {*} = \\frac {1}{M} \\sum_ {j = 1} ^ {M} f (\\hat {r} _ {k j}), \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            181,
+            137,
+            487,
+            181
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nz = \\arg \\max  _ {j} f \\left(\\hat {r} _ {k j}\\right) ^ {\\top} \\cdot \\hat {\\mathbf {r}} _ {k} ^ {*}, \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            191,
+            184,
+            487,
+            209
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {s} = \\frac {f (\\hat {q} _ {k}) + f (\\hat {r} _ {k z})}{2}, \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            191,
+            212,
+            487,
+            244
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $k$  and  $z$  are the finally selected indexes of the rewrite and the response, respectively.",
+        "bbox": [
+            112,
+            253,
+            487,
+            284
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The aggregation for RAR prompting is similar to RTR prompting, but it does not need response selection since there is only one hypothetical response for each rewrite:",
+        "bbox": [
+            112,
+            285,
+            487,
+            349
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {s} = \\frac {f (\\hat {q} _ {k}) + f (\\hat {r} _ {k 1})}{2}. \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            206,
+            359,
+            487,
+            391
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3.3 Mean",
+        "text_level": 1,
+        "bbox": [
+            114,
+            399,
+            220,
+            413
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We average all the rewrite vectors and the corresponding hypothetical response vectors.",
+        "bbox": [
+            112,
+            419,
+            487,
+            450
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "For REW prompting:",
+        "bbox": [
+            131,
+            451,
+            295,
+            467
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {s} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} f \\left(\\hat {q} _ {i}\\right). \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            213,
+            478,
+            487,
+            520
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "For the RTR and RAR prompting methods:",
+        "bbox": [
+            132,
+            531,
+            455,
+            546
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {s} = \\frac {\\sum_ {i = 1} ^ {N} \\left[ f \\left(\\hat {q} _ {i}\\right) + \\sum_ {j = 1} ^ {M} f \\left(\\hat {r} _ {i j}\\right) \\right]}{N * (1 + M)}. \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            154,
+            556,
+            487,
+            596
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Compared with MaxProb and Self-Consistency, the Mean aggregation comprehensively considers more diverse search intent information from all sources. It leverages the collaborative power to enhance the popular intents, but also supplements plausible intents that are missing in a single rewrite or a hypothetical response.",
+        "bbox": [
+            112,
+            605,
+            487,
+            717
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.4 Retrieval",
+        "text_level": 1,
+        "bbox": [
+            112,
+            728,
+            233,
+            741
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "All candidate passages are encoded into passage vectors using the same retriever  $f$ . At search time, we return the passages that are most similar to the final search intent vector  $s$  as the retrieval results.",
+        "bbox": [
+            112,
+            749,
+            487,
+            812
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments",
+        "text_level": 1,
+        "bbox": [
+            112,
+            824,
+            258,
+            841
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Datasets and Metrics",
+        "text_level": 1,
+        "bbox": [
+            112,
+            850,
+            326,
+            864
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We carry out extensive experiments on three widely used conversational search datasets: CAsT-19 (Dalton et al., 2020), CAsT-20 (Dalton et al., 2021), and",
+        "bbox": [
+            112,
+            871,
+            487,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/6ed8a08b7d1547098bbf713a47b1c2f4363b99552412d80df863c743e2f1575d.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td>CAsT-19</td><td>CAsT-20</td><td>CAsT-21</td></tr><tr><td># Conversations</td><td>20</td><td>25</td><td>18</td></tr><tr><td># Turns (Sessions)</td><td>173</td><td>208</td><td>157</td></tr><tr><td># Passages/Docs</td><td colspan=\"2\">38M</td><td>40M</td></tr></table>",
+        "bbox": [
+            514,
+            80,
+            882,
+            153
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Statistics of the three CAsT datasets.",
+        "bbox": [
+            539,
+            162,
+            848,
+            175
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "CAsT-21 (Dalton et al., 2022), which are curated by the human experts of TREC Conversational Assistance Track (CAsT). Each CAsT dataset has dozens of information-seeking conversations comprising hundreds of turns. CAsT-19 and CAsT-20 share the same retrieval corpora while CAsT-21 has a different one. In contrast, CAsT-20 and CAsT-21 have a more complex session structure than CAsT-19 as their questions may refer to previous responses. All three datasets provide human rewrites and passage-level (or document-level) relevance judgments labeled by TREC experts. Table 1 summarizes the basic dataset statistics. $^4$",
+        "bbox": [
+            507,
+            205,
+            882,
+            413
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Following previous work (Dalton et al., 2020, 2021; Yu et al., 2021; Mao et al., 2022a), we adopt Mean Reciprocal Rank (MRR), NDCG@3, and Recall@100 as our evaluation metrics and calculate them using pytrec_eval tool (Van Gysel and de Rijke, 2018). We deem relevance scale  $\\geq 2$  as positive for MRR on CAsT-20 and CAsT-21. For CAsT-21, we split the documents into passages and score each document based on its highest-scored passage (i.e., MaxP (Dai and Callan, 2019)). We conduct the statistical significance tests using paired t-tests at  $p < 0.05$  level.",
+        "bbox": [
+            507,
+            414,
+            882,
+            609
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2 Implementation details",
+        "text_level": 1,
+        "bbox": [
+            507,
+            623,
+            737,
+            638
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We use the OpenAI gpt3.5-turbo-16k as our LLM. The decoding temperature is set to 0.7. We randomly select three conversations from the CAsT225 dataset for demonstration. CAsT-22 is a new conversational search dataset also proposed by TREC CAsT, but only its conversations are released and the relevance judgments have not been made public. Therefore, it cannot be used for evaluation and we just use it for demonstration. For REW prompting, we set  $N = 5$ . For RTR prompting, we set  $N = 1$  and  $M = 5$ . For RAR prompting, we set  $N = 5$ , and  $M$  is naturally set to 1. Following previous studies (Yu et al., 2021; Mao",
+        "bbox": [
+            507,
+            645,
+            882,
+            854
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "4Only the turns that have relevance labels are counted.",
+            "5https://github.com/daltonj/treccastweb/tree/master/2022",
+            "6 Until the submission deadline of EMNLP 2023."
+        ],
+        "bbox": [
+            510,
+            866,
+            873,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "1215",
+        "bbox": [
+            482,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "et al., 2022a,b; Mo et al., 2023a), we adopt the ANCE (Xiong et al., 2021) checkpoint pre-trained on the MSMARCO dataset as our ad-hoc retriever  $f$ . We uniformly truncate the lengths of queries (or rewrites), passages, and hypothetical responses into 64, 256, and 256.",
+        "bbox": [
+            112,
+            84,
+            487,
+            180
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3 Baselines",
+        "text_level": 1,
+        "bbox": [
+            112,
+            191,
+            235,
+            205
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We compare our few-shot LLM4CS against the following six conversational search systems:",
+        "bbox": [
+            112,
+            212,
+            485,
+            243
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "(1) T5QR (Lin et al., 2020): A T5 (Raffel et al., 2020)-based conversational query rewriter trained with the human rewrites.",
+            "(2) ConvDR (Yu et al., 2021): A conversational dense retriever fine-tuned from an ad-hoc retriever by mimicking the representations of human rewrites.",
+            "(3) COTED (Mao et al., 2022a): An improved version of ConvDR (Yu et al., 2021) which incorporates a curriculum learning-based context denoising objective.",
+            "(4) ZeCo (Krasakis et al., 2022): A variant of ColBERT (Khattab and Zaharia, 2020) that matches only the contextualized terms of the current query with passages to perform zero-shot conversational search.",
+            "(5) CRDR (Qian and Dou, 2022): A conversational dense retrieval method where the dense retrieval part is enhanced by the distant supervision from query rewriting in a unified framework.",
+            "(6) ConvGQR (Mo et al., 2023a): A query reformulation framework that combines query rewriting with generative query expansion."
+        ],
+        "bbox": [
+            112,
+            244,
+            487,
+            613
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "T5QR, CRDR, and ConvGQR are trained on the training sessions of QReCC (Anantha et al., 2021), which is a large-scale conversational question answering dataset. The performances of ConvDR and COTED are reported in the few-shot setting using 5-fold cross-validation according to their original papers. We also present the performance of using human rewrites for reference. Note that the same ANCE checkpoint is used to perform dense retrieval for all baselines except ZeCo to ensure fair comparisons.",
+        "bbox": [
+            112,
+            614,
+            487,
+            790
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.4 Main Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            801,
+            265,
+            816
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The overall performance comparisons are presented in Table 2. The reported performance of LLM4CS results from the combination of the RAR prompting method, the Mean aggregation method, and our tailored CoT, which shows to be the most effective combination. We thoroughly investigate the effects",
+        "bbox": [
+            112,
+            822,
+            487,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "of using different prompting and aggregation methods in Section 4.5 and investigate the effects of the incorporation of CoT in Section 4.6.",
+        "bbox": [
+            507,
+            84,
+            882,
+            131
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "From Table 2, we observe that LLM4CS outperforms all the compared baselines in terms of search performance. Specifically, LLM4CS exhibits a relative improvement of over  $18\\%$  compared to the second-best results on the more challenging CAsT-20 and CAsT-21 datasets across all metrics. In particular, even compared to using human rewrites, our LLM4CS can still achieve better results on most metrics, except for the Recall@100 of CAsT-19 and NDCG@3 of CAsT-21. These significant improvements, which are unprecedented in prior research, demonstrate the strong superiority of our LLM4CS over existing methods and underscore the vast potential of using large language models for conversational search.",
+        "bbox": [
+            507,
+            134,
+            884,
+            375
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.5 Effects of Different Prompting Methods and Aggregation Methods",
+        "text_level": 1,
+        "bbox": [
+            507,
+            392,
+            867,
+            423
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We present a comparison of NDCG@3 performance across various prompting and aggregation methods (excluding the incorporation of CoT) in Table 3. Our findings are as follows:",
+        "bbox": [
+            507,
+            432,
+            882,
+            495
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "First, the RAR and RTR prompting methods clearly outperform the REW prompting, demonstrating that the generated hypothetical responses can effectively supplement the short query rewrite to improve retrieval performance. However, even the simple REW prompting can also achieve quite competitive performance compared to existing baselines, particularly on the more challenging CAsT-20 and CAsT-21 datasets, where it shows significant superiority (e.g., 0.380 vs. 0.350 on CAsT-20 and 0.465 vs. 0.385 on CAsT-21). These positive results further highlight the significant advantages of utilizing LLM for conversational search.",
+        "bbox": [
+            507,
+            499,
+            882,
+            708
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Second, in terms of aggregation methods, both Mean and SC consistently outperform MaxProb. These results indicate that depending solely on the top prediction of the language model may not provide sufficient reliability. Instead, utilizing the collective strength of multiple results proves to be a better choice. Additionally, we observe that the Mean aggregation method, which fuses all generated contents into the final search intent vector (Equation 11), does not consistently outperform SC (e.g., on CAsT-20), which actually only fuses one rewrite and one response (Equation 8). This suggests that taking into account more generations",
+        "bbox": [
+            507,
+            709,
+            882,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "1216",
+        "bbox": [
+            482,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/3a9d3be9d6433885d1ff36544f604910fd151b5b7296928b1a22a2dcfab6335f.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">System</td><td colspan=\"3\">CAsT-19</td><td colspan=\"3\">CAsT-20</td><td colspan=\"3\">CAsT-21</td></tr><tr><td>MRR</td><td>NDCG@3</td><td>R@100</td><td>MRR</td><td>NDCG@3</td><td>R@100</td><td>MRR</td><td>NDCG@3</td><td>R@100</td></tr><tr><td colspan=\"10\">Conversational Dense Retrieval</td></tr><tr><td>ConvDR</td><td>0.740</td><td>0.466</td><td>0.362</td><td>0.510</td><td>0.340</td><td>0.345</td><td>0.573</td><td>0.385</td><td>0.483</td></tr><tr><td>COTED</td><td>0.769</td><td>0.478</td><td>0.367</td><td>0.491</td><td>0.342</td><td>0.340</td><td>0.565</td><td>0.371</td><td>0.485</td></tr><tr><td>ZeCo</td><td>-</td><td>0.238‡</td><td>0.216‡</td><td>-</td><td>0.176‡</td><td>0.200‡</td><td>-</td><td>0.234‡</td><td>0.267‡</td></tr><tr><td>CRDR</td><td>0.765</td><td>0.472</td><td>0.357</td><td>0.501</td><td>0.350</td><td>0.313</td><td>0.474</td><td>0.342</td><td>0.380</td></tr></table>",
+        "bbox": [
+            117,
+            80,
+            880,
+            187
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/0d228247fc9b079d5a7b6e249df2bcc563d7690da0b97c0c977df4266092acf6.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"10\">Conversational Query Rewriting</td></tr><tr><td>T5QR</td><td>0.701</td><td>0.417</td><td>0.332</td><td>0.423</td><td>0.299</td><td>0.353</td><td>0.469</td><td>0.330</td><td>0.408</td></tr><tr><td>ConvGQR</td><td>0.708</td><td>0.434</td><td>0.336</td><td>0.465</td><td>0.331</td><td>0.368</td><td>0.433</td><td>0.273</td><td>0.330</td></tr><tr><td>LLM4CS</td><td>0.776†</td><td>0.515†</td><td>0.372†</td><td>0.615†</td><td>0.455†</td><td>0.489†</td><td>0.681†</td><td>0.492†</td><td>0.614†</td></tr><tr><td>Human</td><td>0.740</td><td>0.461</td><td>0.381</td><td>0.591</td><td>0.422</td><td>0.465</td><td>0.680</td><td>0.502</td><td>0.590</td></tr><tr><td>RI-H</td><td>+4.9%</td><td>+11.7%</td><td>-2.4%</td><td>+4.1%</td><td>+7.8%</td><td>+5.2%</td><td>+0.1%</td><td>-2.0%</td><td>+4.1%</td></tr><tr><td>RI-2nd-Best</td><td>+0.9%</td><td>+7.7%</td><td>+1.4%</td><td>+20.6%</td><td>+30.0%</td><td>+32.9%</td><td>+18.8%</td><td>+27.8%</td><td>+26.6%</td></tr></table>",
+        "bbox": [
+            117,
+            189,
+            880,
+            293
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/9f757b3cdd9cb3eac0e8ee93105b5b337610a584142bb7d6eb432f3e15698169.jpg",
+        "table_caption": [
+            "Table 2: Overall performance comparisons.  $\\ddagger$  denotes the results are replicated from their original paper.  $\\dagger$  denotes LLM4CS (RAR + Mean + CoT) significantly outperforms all the compared baselines (except ZeCo) in the  $p < 0.05$  level. The best results are bold and the second-best results are underlined. Human denotes using human rewrites. RI-H and RI-2nd-Best are the relative improvements over Human and the second-best results, respectively."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Aggregation</td><td colspan=\"3\">CAsT-19</td><td colspan=\"3\">CAsT-20</td><td colspan=\"3\">CAsT-21</td></tr><tr><td>REW</td><td>RTR</td><td>RAR</td><td>REW</td><td>RTR</td><td>RAR</td><td>REW</td><td>RTR</td><td>RAR</td></tr><tr><td>MaxProb</td><td>0.441</td><td>0.459</td><td>0.464</td><td>0.356</td><td>0.415</td><td>0.430</td><td>0.407</td><td>0.469</td><td>0.462</td></tr><tr><td>SC</td><td>0.449</td><td>0.466</td><td>0.476</td><td>0.362</td><td>0.432</td><td>0.444</td><td>0.445</td><td>0.473</td><td>0.469</td></tr><tr><td>Mean</td><td>0.447</td><td>0.464</td><td>0.488</td><td>0.380</td><td>0.425</td><td>0.442</td><td>0.465</td><td>0.481</td><td>0.478</td></tr><tr><td>Previous SOTA</td><td></td><td>0.478</td><td></td><td></td><td>0.350</td><td></td><td></td><td>0.385</td><td></td></tr><tr><td>Human</td><td></td><td>0.461</td><td></td><td></td><td>0.422</td><td></td><td></td><td>0.502</td><td></td></tr></table>",
+        "bbox": [
+            121,
+            374,
+            880,
+            489
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Performance comparisons with respect to NDCG@3 using different prompting and aggregation methods. The best combination on each dataset is bold.",
+        "bbox": [
+            112,
+            499,
+            882,
+            527
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/0820c225885095db52f1479dfbd5b5b5110bd7f8de8438cd92dc2378ff9b66c1.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            541,
+            347,
+            639
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/723b533f4804c3c604ea7f07676229a697e61edd123bad6597efac22a7ca7d97.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            381,
+            541,
+            613,
+            638
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/c8c093ce39f7c74bfb4ad8da86a966a6b7947bcf2be09ce5b8860e68e346062b.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            650,
+            542,
+            880,
+            638
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/19c444b61566fd29b5506cd88bbc8092b281d4cbc0f971d0dd78aabaf046c6a5.jpg",
+        "image_caption": [
+            "Figure 2: NDCG@3 comparisons between incorporating our tailored CoT or not across different prompting and aggregation methods on CAsT-20 and CAsT-21 datasets."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            648,
+            346,
+            746
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/7dcd6eda61d96c80080ffed939170e726b07796575622b5a2f40a1842be9fc04.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            383,
+            648,
+            613,
+            745
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/011ce074f1524e9dbf784dda4c90aae13493b6567ff2f0324dc4ea9f4e044908.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            652,
+            649,
+            880,
+            745
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "may not always be beneficial, and a careful selection among them could be helpful to achieve improved results.",
+        "bbox": [
+            112,
+            799,
+            487,
+            846
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.6 Effects of Chain-of-Thought",
+        "text_level": 1,
+        "bbox": [
+            112,
+            863,
+            381,
+            879
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We show the ablation results of our tailored chain-of-thought in Figure 2. We also provide a real",
+        "bbox": [
+            112,
+            887,
+            487,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "example to show how our CoT takes effect in Appendix B.1. From the results, we observe that:",
+        "bbox": [
+            507,
+            799,
+            882,
+            829
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Incorporating our chain-of-thought into all prompting and aggregation methods generally improves search performance. This demonstrates the efficacy of our chain-of-thought in guiding the large language model towards a correct understanding of",
+        "bbox": [
+            507,
+            839,
+            882,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "1217",
+        "bbox": [
+            482,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/4c3f3a922b05f775b36553a4069a881bee9df9039a6075261ff16ff6586caaf7.jpg",
+        "image_caption": [
+            "Figure 3: Human evaluation results for LLM4CS (REW + MaxProb) and T5QR on the three CAsT datasets."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            80,
+            347,
+            180
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/f17eae5810930198ca6bb5e72943ec42405fa14e0c2a63841fdbeb570efb3edc.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            383,
+            80,
+            613,
+            180
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/2ac7e4c79b67e03af3047529f796b80d6c2762ba235893e9f7c4ba9f5580d492.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            650,
+            80,
+            880,
+            180
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "the user's contextual search intent.",
+        "bbox": [
+            112,
+            218,
+            371,
+            231
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In contrast, the improvements are particularly notable for the REW prompting method compared to the RTR and RAR prompting methods. It appears that the introduction of multiple hypothetical responses diminishes the impact of the chain-of-thought. This could be attributed to the fact that including multiple hypothetical responses significantly boosts the quality and robustness of the final search intent vector, thereby reducing the prominence of the chain-of-thought in enhancing search performance.",
+        "bbox": [
+            112,
+            234,
+            489,
+            411
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5 Human Evaluation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            424,
+            314,
+            439
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The retrieval performance is influenced by the ad-hoc retriever used, which implies that automatic search evaluation metrics may not fully reflect the model's capability to understand contextual search intent. Sometimes, two different rewrites can yield significantly different retrieval scores, even though they both accurately represent the user's real search intent. To better investigate the contextual search intent understanding ability of LLM, we perform a fine-grained human evaluation on the rewrites generated by our LLM4CS (REW + MaxProb).",
+        "bbox": [
+            112,
+            451,
+            487,
+            626
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Specifically, we manually compare each model's rewrite with the corresponding human rewrite and label it with one of the following four categories: (1) Good-H: The model's rewrite is nearly the same as the human rewrite. (2) Good-A: The expression of the model's rewrite is different from the human rewrite but it also successfully conveys the user's real search intent. (3) Bad-C: the rewrite has coreference errors. (4) Bad-O: the rewrite omits important contextual information or has other types of errors. Furthermore, we apply the same principle to label the rewrites of T5QR for comparison purposes. A few examples of such categorization are presented in Appendix B.2.",
+        "bbox": [
+            112,
+            629,
+            489,
+            852
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The results of the human evaluation are shown in Figure 3, where we observe that:",
+        "bbox": [
+            112,
+            854,
+            487,
+            885
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "(1) From a human perspective,  $85.5\\%$ ,  $89.4\\%$ , and  $84.8\\%$  of the rewrites of LLM4CS success",
+        "bbox": [
+            112,
+            887,
+            489,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "fully convey the user's real search intent for CAsT-19, CAsT-20, and CAsT-21, respectively. In contrast, the corresponding percentages for T5QR are merely  $75.1\\%$ ,  $62.0\\%$ , and  $58.6\\%$ . Such a high rewriting accuracy of LLM4CS further demonstrates the strong ability of LLM for contextual search intent understanding.",
+        "bbox": [
+            507,
+            217,
+            884,
+            330
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "(2) In the case of CAsT-20 and CAsT-21, a significantly higher percentage of rewrites are labeled as Good-A, in contrast to CAsT-19, where the majority of good rewrites closely resemble the human rewrites. This can be attributed to the higher complexity of the session structure and questions in CAsT-20 and CAsT-21 compared to CAsT-19, which allows for greater freedom in expressing the same search intent.",
+        "bbox": [
+            507,
+            332,
+            884,
+            476
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "(3) The rewrites generated by LLM4CS exhibit coreference errors in less than  $3\\%$  of the cases, whereas T5QR's rewrites contain coreference errors in approximately  $10\\%$  of the cases. This observation highlights the exceptional capability of LLM in addressing coreference issues.",
+        "bbox": [
+            507,
+            480,
+            882,
+            576
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            596,
+            640,
+            612
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we present a simple yet effective prompting framework (i.e., LLM4CS) that leverages LLMs for conversational search. Our framework generates multiple query rewrites and hypothetical responses using tailored prompting methods and aggregates them to robustly represent the user's contextual search intent. Through extensive automatic and human evaluations on three CAsT datasets, we demonstrate its remarkable performance for conversational search. Our study highlights the vast potential of LLMs in conversational search and takes an important initial step in advancing this promising direction. Future research will focus on refining and extending the LLM4CS framework to explore better ways of generation to facilitate search, improving aggregation techniques, optimizing the LLM-retriever interaction, and incorporating reranking strategies.",
+        "bbox": [
+            507,
+            629,
+            885,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "1218",
+        "bbox": [
+            482,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            220,
+            99
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Our work shows that generating multiple rewrites and hypothetical responses and properly aggregating them can effectively improve search performance. However, this requires invoking LLM multiple times, resulting in a higher time cost for retrieval. Due to the relatively high generation latency of LLM, the resulting query latency would be intolerable for users when compared to conventional search engines. A promising approach is to design better prompts capable of obtaining all informative content in one generation, thereby significantly improving query latency. Another limitation is that, similar to the typical disadvantages of CQR methods, the generation process of LLM lacks awareness of the downstream retrieval process. Exploring the utilization of ranking signals to enhance LLM generation would be a compelling direction for future research of conversational search.",
+        "bbox": [
+            115,
+            109,
+            489,
+            397
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgement",
+        "text_level": 1,
+        "bbox": [
+            114,
+            410,
+            278,
+            426
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work was supported by the National Natural Science Foundation of China No. 62272467, Public Computing Cloud, Renmin University of China, and Intelligent Social Governance Platform, Major Innovation & Planning Interdisciplinary Platform for the \"Double-First Class\" Initiative, Renmin University of China, and the Outstanding Innovative Talents Cultivation Funded Programs 2024 of Renmin University of China. The work was partially done at Beijing Key Laboratory of Big Data Management and Analysis Methods.",
+        "bbox": [
+            112,
+            435,
+            489,
+            612
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            639,
+            213,
+            653
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Raviteja Anantha, Svitlana Vakulenko, Zhucheng Tu, Shayne Longpre, Stephen Pulman, and Srinivas Chappidi. 2021. Open-domain question answering goes conversational via question rewriting. In NAACL-HLT, pages 520-534. Association for Computational Linguistics.",
+            "Luiz Henrique Bonifacio, Hugo Abonizio, Marzieh Fadaee, and Rodrigo Frassetto Nogueira. 2022. Inpars: Data augmentation for information retrieval using large language models. CoRR, abs/2202.05144.",
+            "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess,"
+        ],
+        "bbox": [
+            115,
+            661,
+            487,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual.",
+            "Zhiyu Chen, Jie Zhao, Anjie Fang, Besnik Fetahu, Rokhlenko Oleg, and Shervin Malmasi. 2022. Reinforced question rewriting for conversational question answering.",
+            "J Shane Culpepper, Fernando Diaz, and Mark D Smucker. 2018. Research frontiers in information retrieval: Report from the third strategic workshop on information retrieval in lorne (swirl 2018). In ACM SIGIR Forum, volume 52, pages 34-90. ACM New York, NY, USA.",
+            "Zhuyun Dai and Jamie Callan. 2019. Deeper text understanding for IR with contextual neural language modeling. In Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2019, Paris, France, July 21-25, 2019, pages 985-988. ACM.",
+            "Zhuyun Dai, Arun Tejasvi Chaganty, Vincent Y Zhao, Aida Amini, Qazi Mamunur Rashid, Mike Green, and Kelvin Guu. 2022. Dialog inpainting: Turning documents into dialogs. In International Conference on Machine Learning, pages 4558-4586. PMLR.",
+            "Zhuyun Dai, Vincent Y. Zhao, Ji Ma, Yi Luan, Jianmo Ni, Jing Lu, Anton Bakalov, Kelvin Guu, Keith B. Hall, and Ming-Wei Chang. 2023. Promptagator: Few-shot dense retrieval from 8 examples. In 11th International Conference on Learning Representations, ICLR 2023.",
+            "Jeffrey Dalton, Chenyan Xiong, and Jamie Callan. 2020. Trec cast 2019: The conversational assistance track overview. In In Proceedings of TREC.",
+            "Jeffrey Dalton, Chenyan Xiong, and Jamie Callan. 2021. Cast 2020: The conversational assistance track overview. In *In Proceedings of TREC*.",
+            "Jeffrey Dalton, Chenyan Xiong, and Jamie Callan. 2022. Trec cast 2021: The conversational assistance track overview. In In Proceedings of TREC.",
+            "Luyu Gao, Xueguang Ma, Jimmy Lin, and Jamie Callan. 2022. Precise zero-shot dense retrieval without relevance labels. CoRR, abs/2212.10496.",
+            "Vitor Jeronymo, Luiz Henrique Bonifacio, Hugo Abonizio, Marzieh Fadaee, Roberto de Alencar Lotufo, Jakub Zavrel, and Rodrigo Frassetto Nogueira. 2023. Inpars-v2: Large language models as efficient dataset generators for information retrieval. CoRR, abs/2301.01820.",
+            "Dongfu Jiang, Xiang Ren, and Bill Yuchen Lin. 2023. LIm-blender: Ensembling large language models with pairwise ranking and generative fusion. CoRR, abs/2306.02561."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "1219",
+        "bbox": [
+            482,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Omar Khattab and Matei Zaharia. 2020. Colbert: Efficient and effective passage search via contextualized late interaction over BERT. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval (SIGIR), pages 39-48. ACM.",
+            "Sungdong Kim and Gangwoo Kim. 2022. Saving dense retriever from shortcut dependency in conversational search.",
+            "Antonios Minas Krasakis, Andrew Yates, and Evangelos Kanoulas. 2022. Zero-shot query contextualization for conversational search. In Proceedings of the 45th International ACM SIGIR conference on research and development in Information Retrieval (SIGIR).",
+            "Sheng-Chieh Lin, Jheng-Hong Yang, and Jimmy Lin. 2021a. Contextualized query embeddings for conversational search. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP).",
+            "Sheng-Chieh Lin, Jheng-Hong Yang, Rodrigo Nogueira, Ming-Feng Tsai, Chuan-Ju Wang, and Jimmy Lin. 2020. Conversational question reformulation via sequence-to-sequence architectures and pretrained language models. arXiv preprint arXiv:2004.01909.",
+            "Sheng-Chieh Lin, Jheng-Hong Yang, Rodrigo Nogueira, Ming-Feng Tsai, Chuan-Ju Wang, and Jimmy Lin. 2021b. Multi-stage conversational passage retrieval: An approach to fusing term importance estimation and neural query rewriting. ACM Transactions on Information Systems (TOIS), 39(4):1-29.",
+            "Iain Mackie, Shubham Chatterjee, and Jeffrey Dalton. 2023. Generative relevance feedback with large language models. CoRR, abs/2304.13157.",
+            "Kelong Mao, Zhicheng Dou, Bang Liu, Hongjin Qian, Fengran Mo, Xiangli Wu, Xiaohua Cheng, and Zhao Cao. 2023a. Search-oriented conversational query editing. In ACL (Findings), volume ACL 2023 of Findings of ACL. Association for Computational Linguistics.",
+            "Kelong Mao, Zhicheng Dou, and Hongjin Qian. 2022a. Curriculum contrastive context denoising for few-shot conversational dense retrieval. In Proceedings of the 45th International ACM SIGIR conference on research and development in Information Retrieval (SIGIR).",
+            "Kelong Mao, Zhicheng Dou, Hongjin Qian, Fengran Mo, Xiaohua Cheng, and Zhao Cao. 2022b. Convtrans: Transforming web search sessions for conversational dense retrieval. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP).",
+            "Kelong Mao, Hongjin Qian, Fengran Mo, Zhicheng Dou, Bang Liu, Xiaohua Cheng, and Zhao Cao. 2023b. Learning denoised and interpretable session representation for conversational search. In Proceedings of the ACM Web Conference, pages 3193-3202."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Yuning Mao, Pengcheng He, Xiaodong Liu, Yelong Shen, Jianfeng Gao, Jiawei Han, and Weizhu Chen. 2021. Generation-augmented retrieval for open-domain question answering. In ACL/IJCNLP (1), pages 4089-4100. Association for Computational Linguistics.",
+            "Fengran Mo, Kelong Mao, Yutao Zhu, Yihong Wu, Kaiyu Huang, and Jian-Yun Nie. 2023a. ConvGQR: generative query reformulation for conversational search. In ACL, volume ACL 2023. Association for Computational Linguistics.",
+            "Fengran Mo, Jian-Yun Nie, Kaiyu Huang, Kelong Mao, Yutao Zhu, Peng Li, and Yang Liu. 2023b. Learning to relate to previous turns in conversational search. In 29th ACM SIGKDD Conference On Knowledge Discover and Data Mining (SIGKDD).",
+            "Hongjin Qian and Zhicheng Dou. 2022. Explicit query rewriting for conversational dense retrieval. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP).",
+            "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21:140:1-140:67.",
+            "Tao Shen, Guodong Long, Xiubo Geng, Chongyang Tao, Tianyi Zhou, and Daxin Jiang. 2023. Large language models are strong zero-shot retriever. CoRR, abs/2304.14233.",
+            "Weiwei Sun, Lingyong Yan, Xinyu Ma, Pengjie Ren, Dawei Yin, and Zhaochun Ren. 2023. Is chatgpt good at search? investigating large language models as re-ranking agent. CoRR, abs/2304.09542.",
+            "Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Kathleen S. Meier-Hellstern, Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen, Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Aguera y Arcas, Claire Cui, Marian Croak, Ed H. Chi, and Quoc Le. 2022. Lamda: Language models for dialog applications. CoRR, abs/2201.08239.",
+            "Svitlana Vakulenko, Shayne Longpre, Zhucheng Tu, and Raviteja Anantha. 2021a. Question rewriting for conversational question answering. In Proceedings of the 14th ACM International Conference on Web Search and Data Mining (WSDM), pages 355-363."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "1220",
+        "bbox": [
+            482,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Svitlana Vakulenko, Nikos Voskarides, Zhucheng Tu, and Shayne Longpre. 2021b. A comparison of question rewriting methods for conversational passage retrieval. In ECIR (2), volume 12657 of Lecture Notes in Computer Science, pages 418-424. Springer.",
+            "Christophe Van Gysel and Maarten de Rijke. 2018. Pytrec.eval: An extremely fast python interface to trec.eval. In SIGIR. ACM.",
+            "Nikos Voskarides, Dan Li, Pengjie Ren, Evangelos Kanoulas, and Maarten de Rijke. 2020. Query resolution for conversational search with limited supervision. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval (SIGIR), pages 921-930.",
+            "Liang Wang, Nan Yang, and Furu Wei. 2023a. Query2doc: Query expansion with large language models. CoRR, abs/2303.07678.",
+            "Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, and Denny Zhou. 2023b. Self-consistency improves chain of thought reasoning in language models. In 11th International Conference on Learning Representations, ICLR 2023.",
+            "Jason Wei, Maarten Bosma, Vincent Y. Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V. Le. 2022. Finetuned language models are zero-shot learners. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net.",
+            "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed H. Chi, Quoc Le, and Denny Zhou. 2020. Chain of thought prompting elicits reasoning in large language models. Advances in neural information processing systems.",
+            "Zeqiu Wu, Yi Luan, Hannah Rashkin, David Reitter, and Gaurav Singh Tomar. 2022. Conqrr: Conversational query rewriting for retrieval with reinforcement learning.",
+            "Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul N. Bennett, Junaid Ahmed, and Arnold Overwijk. 2021. Approximate nearest neighbor negative contrastive learning for dense text retrieval. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021.",
+            "Shi Yu, Jiahua Liu, Jingqin Yang, Chenyan Xiong, Paul Bennett, Jianfeng Gao, and Zhiyuan Liu. 2020. Few-shot generative conversational query rewriting. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval (SIGIR), pages 1933-1936.",
+            "Shi Yu, Zhenghao Liu, Chenyan Xiong, Tao Feng, and Zhiyuan Liu. 2021. Few-shot conversational dense retrieval. In Proceedings of the 44th International ACM SIGIR conference on research and development in Information Retrieval (SIGIR)."
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Wenhao Yu, Dan Iter, Shuohang Wang, Yichong Xu, Mingxuan Ju, Soumya Sanyal, Chenguang Zhu, Michael Zeng, and Meng Jiang. 2023. Generate rather than retrieve: Large language models are strong context generators. In 11th International Conference on Learning Representations, ICLR 2023.",
+            "Yutao Zhu, Huaying Yuan, Shuting Wang, Jiongnan Liu, Wenhan Liu, Chenlong Deng, Zhicheng Dou, and Ji-Rong Wen. 2023. Large language models for information retrieval: A survey. arXiv preprint arXiv:2308.07107.",
+            "Noah Ziems, Wenhao Yu, Zhihan Zhang, and Meng Jiang. 2023. Large language models are built-in autoregressive search engines. In ACL (Findings), volume ACL 2023 of Findings of ACL. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            316
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "1221",
+        "bbox": [
+            482,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Appendix",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            203,
+            99
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A Prompt of LLM4CS",
+        "text_level": 1,
+        "bbox": [
+            114,
+            108,
+            327,
+            124
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Figure 4 shows a general illustration of the prompt of LLM4CS. The prompt consists of three parts, which are Instruction, Demonstration, and Input. The red part is for REW prompting, the blue part is for the RTR and RAR promptings, and the orange part is for RTR prompting. The green part is for our designed chain-of-thought.",
+        "bbox": [
+            112,
+            135,
+            489,
+            247
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B Case Study",
+        "text_level": 1,
+        "bbox": [
+            114,
+            258,
+            250,
+            275
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B.1 Examples of Chain-of-Thought",
+        "text_level": 1,
+        "bbox": [
+            114,
+            284,
+            408,
+            300
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The example in Table 4 shows how our CoT takes effect. The CoT and Our Rewrite fields are generated by LLM4CS (REW + CoT). We can find that the generated CoT effectively illustrates the rationale behind the rewriting process. Please refer to our anonymously open-sourced repository for more examples.",
+        "bbox": [
+            112,
+            305,
+            489,
+            418
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B.2 Examples of Human Evaluation",
+        "text_level": 1,
+        "bbox": [
+            114,
+            428,
+            413,
+            444
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "An example for each category is shown in Table 5. Please refer to our anonymously open-sourced repository for more examples.",
+        "bbox": [
+            112,
+            449,
+            489,
+            498
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "1222",
+        "bbox": [
+            482,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/d1e9cc80ee8501bea1608e88c6c8bfc53816aa634112229316719d8d47eac500.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Instruction</td><td>For an information-seeking dialog, please help reformulate the question into rewrite that can fully express the user&#x27;s information needs without the need of context, but also generate an informative response to answer the question.</td></tr><tr><td>Demonstration</td><td>I will give you several example multi-turn dialogs, where each turn contains a question, a response, and a rewrite. The rewrite part begins with a sentence explaining the reason for the rewrite.Example #1:Question: What should I consider when buying a phone?Rewrite: This is the first turn. So the question should be rewritten as: What should I consider when buying a phone?Response: The design of the phone and the overall ...Question: Cool. Which one would you recommend?Rewrite: Based on Turn 1, you are inquiring about what should be considered when buying a phone. So the question should be rewritten as: Cool. Which smartphone would you recommend for me?Response: Just because a phone has everything...Example #2:...</td></tr><tr><td>Input</td><td>Your Task (only questions and responses are given):Context:Question: What was the basis of the Watergate scandal?Response: ...Question: ...Response: ...Current Question: So what happened to Nixon?Rewrite: So what happened to Nixon after the events of the Watergate scandal?(Now, you should give me the rewrite and an informative response of the **Current Question** based on the **Context**. The output format should always be: Rewrite:\\$Reason. So the question should be rewritten as: \\$Rewrite\\nResponse: \\$Response. Go ahead!)</td></tr><tr><td>Model Output</td><td>Rewrite: Based on all previous turns, Nixon was badly involved in the Watergate scandal. So the question should be rewritten as: So what happened to Nixon after the events of the Watergate scandal?Response: With the mounting evidence and loss...</td></tr></table>",
+        "bbox": [
+            146,
+            151,
+            853,
+            803
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Figure 4: A general illustration of the prompt of LLM4CS. The prompt consists of three parts, i.e., Instruction, Demonstration, and Input. The red part is for REW prompting, the blue part is for the RTR and RAR prompts, and the orange part is for RTR prompting. The green part is for our designed chain-of-thought.",
+        "bbox": [
+            112,
+            815,
+            884,
+            859
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "1223",
+        "bbox": [
+            482,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/b7042c77c8d16a909151d789af3f4408980f4c57cf99c5a827467116d2bb5b53.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>CAsT-20 (Conversation #84)</td></tr><tr><td>Turn 1</td></tr><tr><td>Question: How does seed investment work for startups?Response: Seed money, sometimes known as seed funding or seed capital, is a form of securities offering in which an investor invests capital in exchange for an equity stake in ... Human Rewrite: How does seed investment work for startups?CoT: This is the first turn.Our Rewrite: How does seed investment work for startups?</td></tr><tr><td>Turn 2</td></tr><tr><td>Question: Tell me more about angel rounds. Response: Tell Me More was a National Public Radio interview show hosted by journalist Michel Martin ... offering sample segments, and soliciting user feedback. It began broadcasting on April 30, 2007. Human Rewrite: Tell me more about angel investment rounds. CoT: Based on the current question and the previous response, we are comparing angel rounds with crowdfunding.Our Rewrite: Tell me more about angel rounds in seed investment for startups.</td></tr><tr><td>Turn 3</td></tr><tr><td>Question: How does it compare with crowdfunding? Response: For crowdfunding of equity stock purchases, there is some research in social psychology that indicates that ... Managing communications with a large number of possibly disappointed investors and supporters can be a substantial, and potentially diverting, task. Human Rewrite: How does an angel investment round compare with crowdfunding? CoT: Based on the current question and the previous response, we are comparing angel rounds with crowdfunding.Our Rewrite: How does angel funding compare to crowdfunding for startups?</td></tr><tr><td>Turn 4</td></tr><tr><td>Question: How much is a typical A series round? Response: OpenGov has raised $47 million from venture capital and angel investors. A $3 million Series A round in 2012 included venture funds 8VC, Founder Collective, Valiant Capital, and a number of high-profile ... Human Rewrite: How much is a typical A series investment round? CoT: Based on the previous turns, we are talking about different types of funding for startups such as seed investment, angel rounds, and crowdfunding.Our Rewrite: How much funding is typically raised in an A series round for startups?</td></tr><tr><td>Turn 5</td></tr><tr><td>Question: What are the risks of investing? Response: Angel investors invest in early stage or start-up companies in exchange for an equity ownership interest. Angel ... Here is what angels particularly care about: The quality, passion, commitment, and integrity of the founders. Human Rewrite: What are the risks of investing in startups? CoT: Based on the previous turns, we are discussing different aspects of startup investments. Our Rewrite: What are the risks associated with investing in startups?</td></tr><tr><td>Turn 6</td></tr><tr><td>Question: What are the different types of instruments used by VCs? Response: ... Human Rewrite: What are the different types of instruments used by VCs? CoT: The user is asking about the different types of instruments used by venture capitalists. Our Rewrite: What are the various types of instruments employed by venture capitalists (VCs)?</td></tr></table>",
+        "bbox": [
+            152,
+            80,
+            842,
+            890
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Table 4: An example to show how our CoT takes effect. The CoT and Our Rewrite fields are generated by LLM4CS (REW + CoT).",
+        "bbox": [
+            112,
+            902,
+            880,
+            929
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "1224",
+        "bbox": [
+            482,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Good-H",
+        "text_level": 1,
+        "bbox": [
+            465,
+            340,
+            531,
+            354
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Question: How about replacing it instead?",
+        "bbox": [
+            161,
+            356,
+            478,
+            372
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Human Rewrite: How much does it cost to replace a garage door opener?",
+        "bbox": [
+            163,
+            373,
+            709,
+            388
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Our Rewrite: What is the cost of replacing a garage door opener?",
+        "bbox": [
+            163,
+            388,
+            648,
+            404
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Good-A",
+        "text_level": 1,
+        "bbox": [
+            465,
+            406,
+            529,
+            418
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Question: What's important for me to know about their safety?",
+        "bbox": [
+            161,
+            420,
+            630,
+            436
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Human Rewrite: What's important for me to know about the safety of smart garage door openers?",
+        "bbox": [
+            161,
+            437,
+            833,
+            468
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Our Rewrite: What are the safety features that I should consider when choosing a garage door opener?",
+        "bbox": [
+            161,
+            469,
+            833,
+            501
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Bad-O",
+        "text_level": 1,
+        "bbox": [
+            470,
+            502,
+            526,
+            514
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Question: Are there any famous foods?",
+        "bbox": [
+            161,
+            518,
+            457,
+            533
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Human Rewrite: Are there any famous foods in Washington D.C.?",
+        "bbox": [
+            163,
+            533,
+            658,
+            550
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Our Rewrite: Are there any famous foods?",
+        "bbox": [
+            163,
+            551,
+            482,
+            565
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Bad-C",
+        "text_level": 1,
+        "bbox": [
+            470,
+            567,
+            524,
+            579
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Question: What is its main economic activity?",
+        "bbox": [
+            161,
+            583,
+            509,
+            598
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Human Rewrite: What is the main economic activity of Salt Lake City?",
+        "bbox": [
+            163,
+            599,
+            695,
+            615
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Our Rewrite: What is the main economic activity in Utah?",
+        "bbox": [
+            163,
+            615,
+            596,
+            630
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Table 5: Examples of the four categories in human evaluation.",
+        "bbox": [
+            287,
+            643,
+            707,
+            657
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "page_number",
+        "text": "1225",
+        "bbox": [
+            482,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 14
+    }
+]

2023/Large Language Models Know Your Contextual Search Intent_ A Prompting Framework for Conversational Search/306f5a35-c68f-4d39-a79b-49d0690590d2_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2378f0fb1d170fb12a4441b33f4ad9314b26b3f89d79f2db28858051a3dc88da
+size 764884

2023/Large Language Models Know Your Contextual Search Intent_ A Prompting Framework for Conversational Search/full.md

@@ -0,0 +1,421 @@
+# Large Language Models Know Your Contextual Search Intent: A Prompting Framework for Conversational Search
+
+Kelong Mao $^{1,2}$ , Zhicheng Dou $^{1,2*}$ , Fengran Mo $^{3}$ , Jiewen Hou $^{4}$ , Haonan Chen $^{1}$ , Hongjin Qian $^{1}$
+
+<sup>1</sup>Gaoling School of Artificial Intelligence, Renmin University of China
+
+$^{2}$ Engineering Research Center of Next-Generation Search and Recommendation, MOE
+
+<sup>3</sup>Université de Montréal, Québec, Canada
+
+$^{4}$ Institute of Computing Technology, Chinese Academy of Sciences
+
+{mk1,dou}@ruc.edu.cn
+
+# Abstract
+
+Precisely understanding users' contextual search intent has been an important challenge for conversational search. As conversational search sessions are much more diverse and long-tailed, existing methods trained on limited data still show unsatisfactory effectiveness and robustness to handle real conversational search scenarios. Recently, large language models (LLMs) have demonstrated amazing capabilities for text generation and conversation understanding. In this work, we present a simple yet effective prompting framework, called LLM4CS, to leverage LLMs as a text-based search intent interpreter to help conversational search. Under this framework, we explore three prompting methods to generate multiple query rewrites and hypothetical responses, and propose to aggregate them into an integrated representation that can robustly represent the user's real contextual search intent. Extensive automatic evaluations and human evaluations on three widely used conversational search benchmarks, including CAsT-19, CAsT-20, and CAsT-21, demonstrate the remarkable performance of our simple LLM4CS framework compared with existing methods and even using human rewrites. Our findings provide important evidence to better understand and leverage LLMs for conversational search. The code is released at https://github.com/kyriemao/LLM4CS.
+
+# 1 Introduction
+
+Conversational search has been expected to be the next generation of search paradigms (Culpepper et al., 2018). It supports search via conversation to provide users with more accurate and intuitive search results and a much more user-friendly search experience. Unlike using traditional search engines which mainly process keyword queries, users could imagine the conversational search system as a knowledgeable human expert and directly start a
+
+multi-turn conversation with it in natural languages to solve their questions. However, one of the main challenges for this beautiful vision is that the users' queries may contain some linguistic problems (e.g., omissions and coreference) and it becomes much harder to capture their real search intent under the multi-turn conversation context (Dalton et al., 2021; Mao et al., 2022a).
+
+To achieve conversational search, an intuitive method known as Conversational Query Rewriting (CQR) involves using a rewriting model to transform the current query into a de-contextualized form. Subsequently, any ad-hoc search models can be seamlessly applied for retrieval purposes. Given that existing ad-hoc search models can be reused directly, CQR demonstrates substantial practical value for industries in quickly initializing their conversational search engines. Another type of method, Conversational Dense Retrieval (CDR), tries to learn a conversational dense retriever to encode the user's real search intent and passages into latent representations and performs dense retrieval. In contrast to the two-step CQR method, where the rewriter is difficult to be directly optimized towards search (Yu et al., 2021; Mao et al., 2023a), the conversational dense retriever can naturally learn from session-passage relevance signals.
+
+However, as conversational search sessions are much more diverse and long-tailed (Mao et al., 2022b; Dai et al., 2022; Mo et al., 2023a), existing CQR and CDR methods trained on limited data still show unsatisfactory performance, especially on more complex conversational search sessions. Many studies (Vakulenko et al., 2021b; Lin et al., 2021a; Qian and Dou, 2022; Krasakis et al., 2022) have demonstrated the performance advantages of using de-contextualized human rewrites on sessions which have complex response dependency. Also, as reported in the public TREC CAst 2021 benchmark (Dalton et al., 2022), existing methods still suffer from significant degradation in their ef
+
+fectiveness as conversations become longer.
+
+Recently, large language models (LLMs) have shown amazing capabilities for text generation and conversation understanding (Brown et al., 2020; Wei et al., 2022; Thoppilan et al., 2022; Zhu et al., 2023). In the field of information retrieval (IR), LLMs have also been successfully utilized to enhance relevance modeling via various techniques such as query generation (Bonifacio et al., 2022; Dai et al., 2023), query expansion (Wang et al., 2023a), document prediction (Gao et al., 2022; Mackie et al., 2023), etc. Inspired by the strong performance of LLMs in conversation and IR, we try to investigate how LLMs can be adapted to precisely grasp users' contextual search intent for conversational search.
+
+In this work, we present a simple yet effective prompting framework, called LLM4CS, to leverage LLM as a search intent interpreter to facilitate conversational search. Specifically, we first prompt LLM to generate both short query rewrites and longer hypothetical responses in multiple perspectives and then aggregate these generated contents into an integrated representation that robustly represents the user's real search intent. Under our framework, we propose three specific prompting methods and aggregation methods, and conduct extensive evaluations on three widely used conversational search benchmarks, including CAsT-19 (Dalton et al., 2020), CAsT-20 (Dalton et al., 2021), and CAsT-21 (Dalton et al., 2022), to comprehensively investigate the effectiveness of LLMs for conversational search.
+
+In general, our framework has two main advantages. First, by leveraging the powerful contextual understanding and generation abilities of large language models, we show that additionally generating hypothetical responses to explicitly supplement more plausible search intents underlying the short rewrite can significantly improve the search performance. Second, we show that properly aggregating multiple rewrites and hypothetical responses can effectively filter out incorrect search intents and enhance the reasonable ones, leading to better search performance and robustness.
+
+Overall, our main contributions are:
+
+- We propose a prompting framework and design three tailored prompting methods to leverage large language models for conversational search, which effectively circumvents the serious data scarcity problem faced by the con
+
+![](images/f990b9bc6f9a08ce86f9fc2c76a6d598b24e489fcc8578691316fe54a36a4f90.jpg)
+
+Context & Question
+
+![](images/8548264f52201c30c17392b3663df4e0506a6f0c858cc5d28141700383dc3db9.jpg)  
+Figure 1: An overview of LLM4CS.
+
+versational search field.
+
+- We show that additionally generating hypothetical responses and properly aggregating multiple generated results are crucial for improving search performance.  
+- We demonstrate the exceptional effectiveness of LLMs for conversational search through both automatic and human evaluations, where the best method in our LLM4CS achieves remarkable improvements in search performance over state-of-the-art CQR and CDR baselines, surpassing even human rewrites.
+
+# 2 Related Work
+
+Conversational Search. Conversational search is an evolving field that involves retrieving relevant information based on multi-turn dialogues with users. To achieve conversational search, two main methods have been developed: conversational query rewriting and conversational dense retrieval. Conversational query rewriting converts the conversational search problem into an ad-hoc search problem by reformulating the search session into a standalone query rewrite. Existing methods try to select useful tokens from the conversation context (Voskarides et al., 2020; Lin et al., 2021b)
+
+or train a generative rewritten based on the pairs of sessions and rewrites (Lin et al., 2020; Yu et al., 2020; Vakulenko et al., 2021a). To make the rewriting process aware of the downstream retrieval process, some studies propose to adopt reinforcement learning (Wu et al., 2022; Chen et al., 2022) or enhance the learning of rewritten with ranking signals (Mao et al., 2023a; Mo et al., 2023a). On the other hand, conversational dense retrieval (Yu et al., 2021) directly encodes the whole conversational search session to perform end-to-end dense retrieval. Existing methods mainly try to improve the session representation through context denoising (Mao et al., 2022a; Krasakis et al., 2022; Mo et al., 2023b; Mao et al., 2023b), data augmentation (Lin et al., 2021a; Mao et al., 2022b; Dai et al., 2022), and hard negative mining (Kim and Kim, 2022).
+
+IR with LLMs. Due to the revolutionary natural language understanding and generation abilities, LLMs are attracting more and more attention from the IR community. LLMs have been leveraged to enhance the relevance modeling of retrieval through query generation (Bonifacio et al., 2022; Jeronymo et al., 2023; Dai et al., 2023), query expansion (Wang et al., 2023a), document prediction (Gao et al., 2022; Mackie et al., 2023), etc. Besides, Shen et al. (2023) proposed to first use the retriever to enhance the generation of LLM and then use the generated content to augment the original search query for better retrieval. Ziems et al. (2023) treated LLM as a built-in search engine to retrieve documents based on the generated URL. There are also some works leveraging LLM to perform re-ranking (Sun et al., 2023; Jiang et al., 2023). Different from previous studies, in this paper, we propose the LLM4CS framework that focuses on studying how LLM can be well utilized to capture the user's contextual search intent to facilitate conversational search.
+
+# 3 LLM4CS: Prompting Large Language Models for Conversational Search
+
+In this section, we introduce our LLM4CS framework, which leverages LLM as a text-based search intent interpreter to facilitate conversational search. Figure 1 shows an overview of LLM4CS. In the following, we first describe our task formulation of conversational search, and then we elaborate on the specific prompting methods and aggregation
+
+methods integrated into the framework. Finally, we introduce the retrieval process.
+
+# 3.1 Task Formulation
+
+We focus on the task of conversational passage retrieval, which is the crucial first step of conversational search that helps the model access the right evidence knowledge. Given the user query  $q^t$  and the conversation context  $C^t = (q^1, r^1, \dots, q^{t-1}, r^{t-1})$  of the current turn  $t$ , where  $q^i$  and  $r^i$  denote the user query and the system response of the historical  $i$ -th turn, our goal is to retrieve passages that are relevant to satisfy the user's real search intent of the current turn.
+
+# 3.2 Prompting Methods
+
+The prompt follows the formulation of [Instruction, Demonstrations, Input], where Input is composed of the query  $q^t$  and the conversation context  $C^t$  of the current turn  $t$ . Figure 4 shows a general illustration of the prompt construction. Specifically, we design and explore three prompting methods, including Rewriting (REW), Rewriting-Then-Response (RTR), and Rewriting-And-Response (RAR), in our LLM4CS framework.
+
+# 3.2.1 Rewriting Prompt (REW)
+
+In this prompting method, we directly treat LLM as a well-trained conversational query rewriter and prompt it to generate rewrites. Only the red part of Figure 4 is enabled. Although straightforward, we show in Section 4.5 that this simple prompting method has been able to achieve quite a strong search performance compared to existing baselines.
+
+# 3.2.2 Rewriting-Then-Response (RTR)
+
+Recently, a few studies (Mao et al., 2021; Gao et al., 2022; Yu et al., 2023; Mackie et al., 2023) have shown that generating hypothetical responses for search queries can often bring positive improvements in retrieval performance. Inspired by them, in addition to prompting LLM to generate rewrites, we continue to utilize the generated rewrites to further prompt LLM to generate hypothetical responses that may contain relevant information to answer the current question. The orange part and the blue part of Figure 4 are enabled. Specifically, we incorporate the pre-generated rewrite (i.e., the orange part) into the Input field of the prompt and
+
+then prompt LLM to generate informative hypothetical responses by referring to the rewrite.
+
+# 3.2.3 Rewriting-And-Response (RAR)
+
+Instead of generating rewrites and hypothetical responses in a two-stage manner, we can also generate them all at once with the red part and the blue part of Figure 4 being enabled. We try to explore whether this one-stage generation could lead to better consistency and accuracy between the generated rewrites and responses, compared with the two-step RTR method.
+
+# 3.2.4 Incorporating Chain-of-Thought
+
+Chain-of-thought (CoT) (Wei et al., 2020) induces the large language models to decompose a reasoning task into multiple intermediate steps which can unlock their stronger reasoning abilities. In this work, we also investigate whether incorporating the chain-of-thought of reasoning the user's real search intent could improve the quality of rewrite and response generation.
+
+Specifically, as shown in the green part of Figure 4, we manually write the chain-of-thought for each turn of the demonstration, which reflects how humans infer the user's real search intent of the current turn based on the historical conversation context. When generating, we instruct LLM to first generate the chain-of-thought before generating rewrites (and responses). We investigate the effects of our proposed CoT tailored to the reasoning of contextual search intent in Section 4.6.
+
+# 3.3 Content Aggregation
+
+After prompting LLM multiple times to generate multiple rewrites and hypothetical responses, we then aggregate these generated contents into an integrated representation to represent the user's complete search intent for search. Let us consider that we have generated  $N$  query rewrites  $\mathcal{Q} = (\hat{q}_1,\dots,\hat{q}_N)$  and  $M$  hypothetical responses  $\mathcal{R} = (\hat{r}_{i1},\dots,\hat{r}_{iM})$  for each rewrite  $\hat{q}_i$ , sorted by their generation probabilities from high to low<sup>2</sup>. Note that in RAR prompting, the rewrites and the hypothetical responses are always generated in pairs (i.e.,  $M = 1$ ). While in RTR prompting, one rewrite can have  $M$  hypothetical responses since they are generated in a two-stage manner. Next, we utilize a dual well-trained ad-hoc retriever<sup>3</sup>  $f$
+
+(e.g, ANCE (Xiong et al., 2021)) to encode each of them into a high-dimensional intent vector and aggregate these intent vectors into one final search intent vector s. Specifically, we design and explore the following three aggregation methods, including MaxProb, Self-Consistency (SC), and Mean, in our LLM4CS framework.
+
+# 3.3.1 MaxProb
+
+We directly use the rewrite and the hypothetical response that have the highest generation probabilities. Therefore, compared with the other two aggregation methods that will be introduced later, MaxProb is highly efficient since it actually does not require multiple generations.
+
+Formally, for REW prompting:
+
+$$
+\mathbf {s} = f \left(\hat {q} _ {1}\right). \tag {1}
+$$
+
+For the RTR and RAR prompting methods, we mix the rewrite and hypothetical response vectors:
+
+$$
+\mathbf {s} = \frac {f (\hat {q} _ {1}) + f (\hat {r} _ {1 1})}{2}. \tag {2}
+$$
+
+# 3.3.2 Self-Consistency (SC)
+
+The multiple generated rewrites and hypothetical responses may express different search intents but only some of them are correct. To obtain a more reasonable and consistent search intent representation, we extend the self-consistency prompting method (Wang et al., 2023b), which was initially designed for reasoning tasks with predetermined answer sets, to our contextual search intent understanding task, which lacks a fixed standard answer. To be specific, we select the intent vector that is the most similar to the cluster center of all intent vectors as the final search intent vector, since it represents the most popular search intent overall.
+
+Formally, for REW prompting:
+
+$$
+\hat {\mathbf {q}} ^ {*} = \frac {1}{N} \sum_ {i = 1} ^ {N} f (\hat {q} _ {i}), \tag {3}
+$$
+
+$$
+\mathbf {s} = \arg \max  _ {f (\hat {q} _ {i})} f (\hat {q} _ {i}) ^ {\top} \cdot \hat {\mathbf {q}} ^ {*}, \tag {4}
+$$
+
+where  $\hat{\mathbf{q}}^*$  is the cluster center vector and  $\cdot$  denotes the dot product that measures the similarity.
+
+For RTR prompting, we first select the intent vector  $f(\hat{q}_k)$  and then select the intent vector  $f(\hat{r}_{kz})$  from all hypothetical responses generated based on
+
+the selected rewrite  $\hat{q}_k$ :
+
+$$
+k = \arg \max _ {i} f (\hat {q} _ {i}) ^ {\top} \cdot \hat {\mathbf {q}} ^ {*}, \tag {5}
+$$
+
+$$
+\hat {\mathbf {r}} _ {k} ^ {*} = \frac {1}{M} \sum_ {j = 1} ^ {M} f (\hat {r} _ {k j}), \tag {6}
+$$
+
+$$
+z = \arg \max  _ {j} f \left(\hat {r} _ {k j}\right) ^ {\top} \cdot \hat {\mathbf {r}} _ {k} ^ {*}, \tag {7}
+$$
+
+$$
+\mathbf {s} = \frac {f (\hat {q} _ {k}) + f (\hat {r} _ {k z})}{2}, \tag {8}
+$$
+
+where  $k$  and  $z$  are the finally selected indexes of the rewrite and the response, respectively.
+
+The aggregation for RAR prompting is similar to RTR prompting, but it does not need response selection since there is only one hypothetical response for each rewrite:
+
+$$
+\mathbf {s} = \frac {f (\hat {q} _ {k}) + f (\hat {r} _ {k 1})}{2}. \tag {9}
+$$
+
+# 3.3.3 Mean
+
+We average all the rewrite vectors and the corresponding hypothetical response vectors.
+
+For REW prompting:
+
+$$
+\mathbf {s} = \frac {1}{N} \sum_ {i = 1} ^ {N} f \left(\hat {q} _ {i}\right). \tag {10}
+$$
+
+For the RTR and RAR prompting methods:
+
+$$
+\mathbf {s} = \frac {\sum_ {i = 1} ^ {N} \left[ f \left(\hat {q} _ {i}\right) + \sum_ {j = 1} ^ {M} f \left(\hat {r} _ {i j}\right) \right]}{N * (1 + M)}. \tag {11}
+$$
+
+Compared with MaxProb and Self-Consistency, the Mean aggregation comprehensively considers more diverse search intent information from all sources. It leverages the collaborative power to enhance the popular intents, but also supplements plausible intents that are missing in a single rewrite or a hypothetical response.
+
+# 3.4 Retrieval
+
+All candidate passages are encoded into passage vectors using the same retriever  $f$ . At search time, we return the passages that are most similar to the final search intent vector  $s$  as the retrieval results.
+
+# 4 Experiments
+
+# 4.1 Datasets and Metrics
+
+We carry out extensive experiments on three widely used conversational search datasets: CAsT-19 (Dalton et al., 2020), CAsT-20 (Dalton et al., 2021), and
+
+<table><tr><td>Dataset</td><td>CAsT-19</td><td>CAsT-20</td><td>CAsT-21</td></tr><tr><td># Conversations</td><td>20</td><td>25</td><td>18</td></tr><tr><td># Turns (Sessions)</td><td>173</td><td>208</td><td>157</td></tr><tr><td># Passages/Docs</td><td colspan="2">38M</td><td>40M</td></tr></table>
+
+Table 1: Statistics of the three CAsT datasets.
+
+CAsT-21 (Dalton et al., 2022), which are curated by the human experts of TREC Conversational Assistance Track (CAsT). Each CAsT dataset has dozens of information-seeking conversations comprising hundreds of turns. CAsT-19 and CAsT-20 share the same retrieval corpora while CAsT-21 has a different one. In contrast, CAsT-20 and CAsT-21 have a more complex session structure than CAsT-19 as their questions may refer to previous responses. All three datasets provide human rewrites and passage-level (or document-level) relevance judgments labeled by TREC experts. Table 1 summarizes the basic dataset statistics. $^4$
+
+Following previous work (Dalton et al., 2020, 2021; Yu et al., 2021; Mao et al., 2022a), we adopt Mean Reciprocal Rank (MRR), NDCG@3, and Recall@100 as our evaluation metrics and calculate them using pytrec_eval tool (Van Gysel and de Rijke, 2018). We deem relevance scale  $\geq 2$  as positive for MRR on CAsT-20 and CAsT-21. For CAsT-21, we split the documents into passages and score each document based on its highest-scored passage (i.e., MaxP (Dai and Callan, 2019)). We conduct the statistical significance tests using paired t-tests at  $p < 0.05$  level.
+
+# 4.2 Implementation details
+
+We use the OpenAI gpt3.5-turbo-16k as our LLM. The decoding temperature is set to 0.7. We randomly select three conversations from the CAsT225 dataset for demonstration. CAsT-22 is a new conversational search dataset also proposed by TREC CAsT, but only its conversations are released and the relevance judgments have not been made public. Therefore, it cannot be used for evaluation and we just use it for demonstration. For REW prompting, we set  $N = 5$ . For RTR prompting, we set  $N = 1$  and  $M = 5$ . For RAR prompting, we set  $N = 5$ , and  $M$  is naturally set to 1. Following previous studies (Yu et al., 2021; Mao
+
+4Only the turns that have relevance labels are counted.  
+5https://github.com/daltonj/treccastweb/tree/master/2022  
+6 Until the submission deadline of EMNLP 2023.
+
+et al., 2022a,b; Mo et al., 2023a), we adopt the ANCE (Xiong et al., 2021) checkpoint pre-trained on the MSMARCO dataset as our ad-hoc retriever  $f$ . We uniformly truncate the lengths of queries (or rewrites), passages, and hypothetical responses into 64, 256, and 256.
+
+# 4.3 Baselines
+
+We compare our few-shot LLM4CS against the following six conversational search systems:
+
+(1) T5QR (Lin et al., 2020): A T5 (Raffel et al., 2020)-based conversational query rewriter trained with the human rewrites.  
+(2) ConvDR (Yu et al., 2021): A conversational dense retriever fine-tuned from an ad-hoc retriever by mimicking the representations of human rewrites.  
+(3) COTED (Mao et al., 2022a): An improved version of ConvDR (Yu et al., 2021) which incorporates a curriculum learning-based context denoising objective.  
+(4) ZeCo (Krasakis et al., 2022): A variant of ColBERT (Khattab and Zaharia, 2020) that matches only the contextualized terms of the current query with passages to perform zero-shot conversational search.  
+(5) CRDR (Qian and Dou, 2022): A conversational dense retrieval method where the dense retrieval part is enhanced by the distant supervision from query rewriting in a unified framework.  
+(6) ConvGQR (Mo et al., 2023a): A query reformulation framework that combines query rewriting with generative query expansion.
+
+T5QR, CRDR, and ConvGQR are trained on the training sessions of QReCC (Anantha et al., 2021), which is a large-scale conversational question answering dataset. The performances of ConvDR and COTED are reported in the few-shot setting using 5-fold cross-validation according to their original papers. We also present the performance of using human rewrites for reference. Note that the same ANCE checkpoint is used to perform dense retrieval for all baselines except ZeCo to ensure fair comparisons.
+
+# 4.4 Main Results
+
+The overall performance comparisons are presented in Table 2. The reported performance of LLM4CS results from the combination of the RAR prompting method, the Mean aggregation method, and our tailored CoT, which shows to be the most effective combination. We thoroughly investigate the effects
+
+of using different prompting and aggregation methods in Section 4.5 and investigate the effects of the incorporation of CoT in Section 4.6.
+
+From Table 2, we observe that LLM4CS outperforms all the compared baselines in terms of search performance. Specifically, LLM4CS exhibits a relative improvement of over  $18\%$  compared to the second-best results on the more challenging CAsT-20 and CAsT-21 datasets across all metrics. In particular, even compared to using human rewrites, our LLM4CS can still achieve better results on most metrics, except for the Recall@100 of CAsT-19 and NDCG@3 of CAsT-21. These significant improvements, which are unprecedented in prior research, demonstrate the strong superiority of our LLM4CS over existing methods and underscore the vast potential of using large language models for conversational search.
+
+# 4.5 Effects of Different Prompting Methods and Aggregation Methods
+
+We present a comparison of NDCG@3 performance across various prompting and aggregation methods (excluding the incorporation of CoT) in Table 3. Our findings are as follows:
+
+First, the RAR and RTR prompting methods clearly outperform the REW prompting, demonstrating that the generated hypothetical responses can effectively supplement the short query rewrite to improve retrieval performance. However, even the simple REW prompting can also achieve quite competitive performance compared to existing baselines, particularly on the more challenging CAsT-20 and CAsT-21 datasets, where it shows significant superiority (e.g., 0.380 vs. 0.350 on CAsT-20 and 0.465 vs. 0.385 on CAsT-21). These positive results further highlight the significant advantages of utilizing LLM for conversational search.
+
+Second, in terms of aggregation methods, both Mean and SC consistently outperform MaxProb. These results indicate that depending solely on the top prediction of the language model may not provide sufficient reliability. Instead, utilizing the collective strength of multiple results proves to be a better choice. Additionally, we observe that the Mean aggregation method, which fuses all generated contents into the final search intent vector (Equation 11), does not consistently outperform SC (e.g., on CAsT-20), which actually only fuses one rewrite and one response (Equation 8). This suggests that taking into account more generations
+
+<table><tr><td rowspan="2">System</td><td colspan="3">CAsT-19</td><td colspan="3">CAsT-20</td><td colspan="3">CAsT-21</td></tr><tr><td>MRR</td><td>NDCG@3</td><td>R@100</td><td>MRR</td><td>NDCG@3</td><td>R@100</td><td>MRR</td><td>NDCG@3</td><td>R@100</td></tr><tr><td colspan="10">Conversational Dense Retrieval</td></tr><tr><td>ConvDR</td><td>0.740</td><td>0.466</td><td>0.362</td><td>0.510</td><td>0.340</td><td>0.345</td><td>0.573</td><td>0.385</td><td>0.483</td></tr><tr><td>COTED</td><td>0.769</td><td>0.478</td><td>0.367</td><td>0.491</td><td>0.342</td><td>0.340</td><td>0.565</td><td>0.371</td><td>0.485</td></tr><tr><td>ZeCo</td><td>-</td><td>0.238‡</td><td>0.216‡</td><td>-</td><td>0.176‡</td><td>0.200‡</td><td>-</td><td>0.234‡</td><td>0.267‡</td></tr><tr><td>CRDR</td><td>0.765</td><td>0.472</td><td>0.357</td><td>0.501</td><td>0.350</td><td>0.313</td><td>0.474</td><td>0.342</td><td>0.380</td></tr></table>
+
+<table><tr><td colspan="10">Conversational Query Rewriting</td></tr><tr><td>T5QR</td><td>0.701</td><td>0.417</td><td>0.332</td><td>0.423</td><td>0.299</td><td>0.353</td><td>0.469</td><td>0.330</td><td>0.408</td></tr><tr><td>ConvGQR</td><td>0.708</td><td>0.434</td><td>0.336</td><td>0.465</td><td>0.331</td><td>0.368</td><td>0.433</td><td>0.273</td><td>0.330</td></tr><tr><td>LLM4CS</td><td>0.776†</td><td>0.515†</td><td>0.372†</td><td>0.615†</td><td>0.455†</td><td>0.489†</td><td>0.681†</td><td>0.492†</td><td>0.614†</td></tr><tr><td>Human</td><td>0.740</td><td>0.461</td><td>0.381</td><td>0.591</td><td>0.422</td><td>0.465</td><td>0.680</td><td>0.502</td><td>0.590</td></tr><tr><td>RI-H</td><td>+4.9%</td><td>+11.7%</td><td>-2.4%</td><td>+4.1%</td><td>+7.8%</td><td>+5.2%</td><td>+0.1%</td><td>-2.0%</td><td>+4.1%</td></tr><tr><td>RI-2nd-Best</td><td>+0.9%</td><td>+7.7%</td><td>+1.4%</td><td>+20.6%</td><td>+30.0%</td><td>+32.9%</td><td>+18.8%</td><td>+27.8%</td><td>+26.6%</td></tr></table>
+
+Table 2: Overall performance comparisons.  $\ddagger$  denotes the results are replicated from their original paper.  $\dagger$  denotes LLM4CS (RAR + Mean + CoT) significantly outperforms all the compared baselines (except ZeCo) in the  $p < 0.05$  level. The best results are bold and the second-best results are underlined. Human denotes using human rewrites. RI-H and RI-2nd-Best are the relative improvements over Human and the second-best results, respectively.  
+
+<table><tr><td rowspan="2">Aggregation</td><td colspan="3">CAsT-19</td><td colspan="3">CAsT-20</td><td colspan="3">CAsT-21</td></tr><tr><td>REW</td><td>RTR</td><td>RAR</td><td>REW</td><td>RTR</td><td>RAR</td><td>REW</td><td>RTR</td><td>RAR</td></tr><tr><td>MaxProb</td><td>0.441</td><td>0.459</td><td>0.464</td><td>0.356</td><td>0.415</td><td>0.430</td><td>0.407</td><td>0.469</td><td>0.462</td></tr><tr><td>SC</td><td>0.449</td><td>0.466</td><td>0.476</td><td>0.362</td><td>0.432</td><td>0.444</td><td>0.445</td><td>0.473</td><td>0.469</td></tr><tr><td>Mean</td><td>0.447</td><td>0.464</td><td>0.488</td><td>0.380</td><td>0.425</td><td>0.442</td><td>0.465</td><td>0.481</td><td>0.478</td></tr><tr><td>Previous SOTA</td><td></td><td>0.478</td><td></td><td></td><td>0.350</td><td></td><td></td><td>0.385</td><td></td></tr><tr><td>Human</td><td></td><td>0.461</td><td></td><td></td><td>0.422</td><td></td><td></td><td>0.502</td><td></td></tr></table>
+
+Table 3: Performance comparisons with respect to NDCG@3 using different prompting and aggregation methods. The best combination on each dataset is bold.
+
+![](images/0820c225885095db52f1479dfbd5b5b5110bd7f8de8438cd92dc2378ff9b66c1.jpg)
+
+![](images/723b533f4804c3c604ea7f07676229a697e61edd123bad6597efac22a7ca7d97.jpg)
+
+![](images/c8c093ce39f7c74bfb4ad8da86a966a6b7947bcf2be09ce5b8860e68e346062b.jpg)
+
+![](images/19c444b61566fd29b5506cd88bbc8092b281d4cbc0f971d0dd78aabaf046c6a5.jpg)  
+Figure 2: NDCG@3 comparisons between incorporating our tailored CoT or not across different prompting and aggregation methods on CAsT-20 and CAsT-21 datasets.
+
+![](images/7dcd6eda61d96c80080ffed939170e726b07796575622b5a2f40a1842be9fc04.jpg)
+
+![](images/011ce074f1524e9dbf784dda4c90aae13493b6567ff2f0324dc4ea9f4e044908.jpg)
+
+may not always be beneficial, and a careful selection among them could be helpful to achieve improved results.
+
+# 4.6 Effects of Chain-of-Thought
+
+We show the ablation results of our tailored chain-of-thought in Figure 2. We also provide a real
+
+example to show how our CoT takes effect in Appendix B.1. From the results, we observe that:
+
+Incorporating our chain-of-thought into all prompting and aggregation methods generally improves search performance. This demonstrates the efficacy of our chain-of-thought in guiding the large language model towards a correct understanding of
+
+![](images/4c3f3a922b05f775b36553a4069a881bee9df9039a6075261ff16ff6586caaf7.jpg)  
+Figure 3: Human evaluation results for LLM4CS (REW + MaxProb) and T5QR on the three CAsT datasets.
+
+![](images/f17eae5810930198ca6bb5e72943ec42405fa14e0c2a63841fdbeb570efb3edc.jpg)
+
+![](images/2ac7e4c79b67e03af3047529f796b80d6c2762ba235893e9f7c4ba9f5580d492.jpg)
+
+the user's contextual search intent.
+
+In contrast, the improvements are particularly notable for the REW prompting method compared to the RTR and RAR prompting methods. It appears that the introduction of multiple hypothetical responses diminishes the impact of the chain-of-thought. This could be attributed to the fact that including multiple hypothetical responses significantly boosts the quality and robustness of the final search intent vector, thereby reducing the prominence of the chain-of-thought in enhancing search performance.
+
+# 5 Human Evaluation
+
+The retrieval performance is influenced by the ad-hoc retriever used, which implies that automatic search evaluation metrics may not fully reflect the model's capability to understand contextual search intent. Sometimes, two different rewrites can yield significantly different retrieval scores, even though they both accurately represent the user's real search intent. To better investigate the contextual search intent understanding ability of LLM, we perform a fine-grained human evaluation on the rewrites generated by our LLM4CS (REW + MaxProb).
+
+Specifically, we manually compare each model's rewrite with the corresponding human rewrite and label it with one of the following four categories: (1) Good-H: The model's rewrite is nearly the same as the human rewrite. (2) Good-A: The expression of the model's rewrite is different from the human rewrite but it also successfully conveys the user's real search intent. (3) Bad-C: the rewrite has coreference errors. (4) Bad-O: the rewrite omits important contextual information or has other types of errors. Furthermore, we apply the same principle to label the rewrites of T5QR for comparison purposes. A few examples of such categorization are presented in Appendix B.2.
+
+The results of the human evaluation are shown in Figure 3, where we observe that:
+
+(1) From a human perspective,  $85.5\%$ ,  $89.4\%$ , and  $84.8\%$  of the rewrites of LLM4CS success
+
+fully convey the user's real search intent for CAsT-19, CAsT-20, and CAsT-21, respectively. In contrast, the corresponding percentages for T5QR are merely  $75.1\%$ ,  $62.0\%$ , and  $58.6\%$ . Such a high rewriting accuracy of LLM4CS further demonstrates the strong ability of LLM for contextual search intent understanding.
+
+(2) In the case of CAsT-20 and CAsT-21, a significantly higher percentage of rewrites are labeled as Good-A, in contrast to CAsT-19, where the majority of good rewrites closely resemble the human rewrites. This can be attributed to the higher complexity of the session structure and questions in CAsT-20 and CAsT-21 compared to CAsT-19, which allows for greater freedom in expressing the same search intent.
+
+(3) The rewrites generated by LLM4CS exhibit coreference errors in less than  $3\%$  of the cases, whereas T5QR's rewrites contain coreference errors in approximately  $10\%$  of the cases. This observation highlights the exceptional capability of LLM in addressing coreference issues.
+
+# 6 Conclusion
+
+In this paper, we present a simple yet effective prompting framework (i.e., LLM4CS) that leverages LLMs for conversational search. Our framework generates multiple query rewrites and hypothetical responses using tailored prompting methods and aggregates them to robustly represent the user's contextual search intent. Through extensive automatic and human evaluations on three CAsT datasets, we demonstrate its remarkable performance for conversational search. Our study highlights the vast potential of LLMs in conversational search and takes an important initial step in advancing this promising direction. Future research will focus on refining and extending the LLM4CS framework to explore better ways of generation to facilitate search, improving aggregation techniques, optimizing the LLM-retriever interaction, and incorporating reranking strategies.
+
+# Limitations
+
+Our work shows that generating multiple rewrites and hypothetical responses and properly aggregating them can effectively improve search performance. However, this requires invoking LLM multiple times, resulting in a higher time cost for retrieval. Due to the relatively high generation latency of LLM, the resulting query latency would be intolerable for users when compared to conventional search engines. A promising approach is to design better prompts capable of obtaining all informative content in one generation, thereby significantly improving query latency. Another limitation is that, similar to the typical disadvantages of CQR methods, the generation process of LLM lacks awareness of the downstream retrieval process. Exploring the utilization of ranking signals to enhance LLM generation would be a compelling direction for future research of conversational search.
+
+# Acknowledgement
+
+This work was supported by the National Natural Science Foundation of China No. 62272467, Public Computing Cloud, Renmin University of China, and Intelligent Social Governance Platform, Major Innovation & Planning Interdisciplinary Platform for the "Double-First Class" Initiative, Renmin University of China, and the Outstanding Innovative Talents Cultivation Funded Programs 2024 of Renmin University of China. The work was partially done at Beijing Key Laboratory of Big Data Management and Analysis Methods.
+
+# References
+
+Raviteja Anantha, Svitlana Vakulenko, Zhucheng Tu, Shayne Longpre, Stephen Pulman, and Srinivas Chappidi. 2021. Open-domain question answering goes conversational via question rewriting. In NAACL-HLT, pages 520-534. Association for Computational Linguistics.  
+Luiz Henrique Bonifacio, Hugo Abonizio, Marzieh Fadaee, and Rodrigo Frassetto Nogueira. 2022. Inpars: Data augmentation for information retrieval using large language models. CoRR, abs/2202.05144.  
+Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess,
+
+Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual.  
+Zhiyu Chen, Jie Zhao, Anjie Fang, Besnik Fetahu, Rokhlenko Oleg, and Shervin Malmasi. 2022. Reinforced question rewriting for conversational question answering.  
+J Shane Culpepper, Fernando Diaz, and Mark D Smucker. 2018. Research frontiers in information retrieval: Report from the third strategic workshop on information retrieval in lorne (swirl 2018). In ACM SIGIR Forum, volume 52, pages 34-90. ACM New York, NY, USA.  
+Zhuyun Dai and Jamie Callan. 2019. Deeper text understanding for IR with contextual neural language modeling. In Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2019, Paris, France, July 21-25, 2019, pages 985-988. ACM.  
+Zhuyun Dai, Arun Tejasvi Chaganty, Vincent Y Zhao, Aida Amini, Qazi Mamunur Rashid, Mike Green, and Kelvin Guu. 2022. Dialog inpainting: Turning documents into dialogs. In International Conference on Machine Learning, pages 4558-4586. PMLR.  
+Zhuyun Dai, Vincent Y. Zhao, Ji Ma, Yi Luan, Jianmo Ni, Jing Lu, Anton Bakalov, Kelvin Guu, Keith B. Hall, and Ming-Wei Chang. 2023. Promptagator: Few-shot dense retrieval from 8 examples. In 11th International Conference on Learning Representations, ICLR 2023.  
+Jeffrey Dalton, Chenyan Xiong, and Jamie Callan. 2020. Trec cast 2019: The conversational assistance track overview. In In Proceedings of TREC.  
+Jeffrey Dalton, Chenyan Xiong, and Jamie Callan. 2021. Cast 2020: The conversational assistance track overview. In *In Proceedings of TREC*.  
+Jeffrey Dalton, Chenyan Xiong, and Jamie Callan. 2022. Trec cast 2021: The conversational assistance track overview. In In Proceedings of TREC.  
+Luyu Gao, Xueguang Ma, Jimmy Lin, and Jamie Callan. 2022. Precise zero-shot dense retrieval without relevance labels. CoRR, abs/2212.10496.  
+Vitor Jeronymo, Luiz Henrique Bonifacio, Hugo Abonizio, Marzieh Fadaee, Roberto de Alencar Lotufo, Jakub Zavrel, and Rodrigo Frassetto Nogueira. 2023. Inpars-v2: Large language models as efficient dataset generators for information retrieval. CoRR, abs/2301.01820.  
+Dongfu Jiang, Xiang Ren, and Bill Yuchen Lin. 2023. LIm-blender: Ensembling large language models with pairwise ranking and generative fusion. CoRR, abs/2306.02561.
+
+Omar Khattab and Matei Zaharia. 2020. Colbert: Efficient and effective passage search via contextualized late interaction over BERT. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval (SIGIR), pages 39-48. ACM.  
+Sungdong Kim and Gangwoo Kim. 2022. Saving dense retriever from shortcut dependency in conversational search.  
+Antonios Minas Krasakis, Andrew Yates, and Evangelos Kanoulas. 2022. Zero-shot query contextualization for conversational search. In Proceedings of the 45th International ACM SIGIR conference on research and development in Information Retrieval (SIGIR).  
+Sheng-Chieh Lin, Jheng-Hong Yang, and Jimmy Lin. 2021a. Contextualized query embeddings for conversational search. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP).  
+Sheng-Chieh Lin, Jheng-Hong Yang, Rodrigo Nogueira, Ming-Feng Tsai, Chuan-Ju Wang, and Jimmy Lin. 2020. Conversational question reformulation via sequence-to-sequence architectures and pretrained language models. arXiv preprint arXiv:2004.01909.  
+Sheng-Chieh Lin, Jheng-Hong Yang, Rodrigo Nogueira, Ming-Feng Tsai, Chuan-Ju Wang, and Jimmy Lin. 2021b. Multi-stage conversational passage retrieval: An approach to fusing term importance estimation and neural query rewriting. ACM Transactions on Information Systems (TOIS), 39(4):1-29.  
+Iain Mackie, Shubham Chatterjee, and Jeffrey Dalton. 2023. Generative relevance feedback with large language models. CoRR, abs/2304.13157.  
+Kelong Mao, Zhicheng Dou, Bang Liu, Hongjin Qian, Fengran Mo, Xiangli Wu, Xiaohua Cheng, and Zhao Cao. 2023a. Search-oriented conversational query editing. In ACL (Findings), volume ACL 2023 of Findings of ACL. Association for Computational Linguistics.  
+Kelong Mao, Zhicheng Dou, and Hongjin Qian. 2022a. Curriculum contrastive context denoising for few-shot conversational dense retrieval. In Proceedings of the 45th International ACM SIGIR conference on research and development in Information Retrieval (SIGIR).  
+Kelong Mao, Zhicheng Dou, Hongjin Qian, Fengran Mo, Xiaohua Cheng, and Zhao Cao. 2022b. Convtrans: Transforming web search sessions for conversational dense retrieval. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP).  
+Kelong Mao, Hongjin Qian, Fengran Mo, Zhicheng Dou, Bang Liu, Xiaohua Cheng, and Zhao Cao. 2023b. Learning denoised and interpretable session representation for conversational search. In Proceedings of the ACM Web Conference, pages 3193-3202.
+
+Yuning Mao, Pengcheng He, Xiaodong Liu, Yelong Shen, Jianfeng Gao, Jiawei Han, and Weizhu Chen. 2021. Generation-augmented retrieval for open-domain question answering. In ACL/IJCNLP (1), pages 4089-4100. Association for Computational Linguistics.  
+Fengran Mo, Kelong Mao, Yutao Zhu, Yihong Wu, Kaiyu Huang, and Jian-Yun Nie. 2023a. ConvGQR: generative query reformulation for conversational search. In ACL, volume ACL 2023. Association for Computational Linguistics.  
+Fengran Mo, Jian-Yun Nie, Kaiyu Huang, Kelong Mao, Yutao Zhu, Peng Li, and Yang Liu. 2023b. Learning to relate to previous turns in conversational search. In 29th ACM SIGKDD Conference On Knowledge Discover and Data Mining (SIGKDD).  
+Hongjin Qian and Zhicheng Dou. 2022. Explicit query rewriting for conversational dense retrieval. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP).  
+Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21:140:1-140:67.  
+Tao Shen, Guodong Long, Xiubo Geng, Chongyang Tao, Tianyi Zhou, and Daxin Jiang. 2023. Large language models are strong zero-shot retriever. CoRR, abs/2304.14233.  
+Weiwei Sun, Lingyong Yan, Xinyu Ma, Pengjie Ren, Dawei Yin, and Zhaochun Ren. 2023. Is chatgpt good at search? investigating large language models as re-ranking agent. CoRR, abs/2304.09542.  
+Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Kathleen S. Meier-Hellstern, Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen, Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Aguera y Arcas, Claire Cui, Marian Croak, Ed H. Chi, and Quoc Le. 2022. Lamda: Language models for dialog applications. CoRR, abs/2201.08239.  
+Svitlana Vakulenko, Shayne Longpre, Zhucheng Tu, and Raviteja Anantha. 2021a. Question rewriting for conversational question answering. In Proceedings of the 14th ACM International Conference on Web Search and Data Mining (WSDM), pages 355-363.
+
+Svitlana Vakulenko, Nikos Voskarides, Zhucheng Tu, and Shayne Longpre. 2021b. A comparison of question rewriting methods for conversational passage retrieval. In ECIR (2), volume 12657 of Lecture Notes in Computer Science, pages 418-424. Springer.  
+Christophe Van Gysel and Maarten de Rijke. 2018. Pytrec.eval: An extremely fast python interface to trec.eval. In SIGIR. ACM.  
+Nikos Voskarides, Dan Li, Pengjie Ren, Evangelos Kanoulas, and Maarten de Rijke. 2020. Query resolution for conversational search with limited supervision. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval (SIGIR), pages 921-930.  
+Liang Wang, Nan Yang, and Furu Wei. 2023a. Query2doc: Query expansion with large language models. CoRR, abs/2303.07678.  
+Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, and Denny Zhou. 2023b. Self-consistency improves chain of thought reasoning in language models. In 11th International Conference on Learning Representations, ICLR 2023.  
+Jason Wei, Maarten Bosma, Vincent Y. Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V. Le. 2022. Finetuned language models are zero-shot learners. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net.  
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed H. Chi, Quoc Le, and Denny Zhou. 2020. Chain of thought prompting elicits reasoning in large language models. Advances in neural information processing systems.  
+Zeqiu Wu, Yi Luan, Hannah Rashkin, David Reitter, and Gaurav Singh Tomar. 2022. Conqrr: Conversational query rewriting for retrieval with reinforcement learning.  
+Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul N. Bennett, Junaid Ahmed, and Arnold Overwijk. 2021. Approximate nearest neighbor negative contrastive learning for dense text retrieval. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021.  
+Shi Yu, Jiahua Liu, Jingqin Yang, Chenyan Xiong, Paul Bennett, Jianfeng Gao, and Zhiyuan Liu. 2020. Few-shot generative conversational query rewriting. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval (SIGIR), pages 1933-1936.  
+Shi Yu, Zhenghao Liu, Chenyan Xiong, Tao Feng, and Zhiyuan Liu. 2021. Few-shot conversational dense retrieval. In Proceedings of the 44th International ACM SIGIR conference on research and development in Information Retrieval (SIGIR).
+
+Wenhao Yu, Dan Iter, Shuohang Wang, Yichong Xu, Mingxuan Ju, Soumya Sanyal, Chenguang Zhu, Michael Zeng, and Meng Jiang. 2023. Generate rather than retrieve: Large language models are strong context generators. In 11th International Conference on Learning Representations, ICLR 2023.  
+Yutao Zhu, Huaying Yuan, Shuting Wang, Jiongnan Liu, Wenhan Liu, Chenlong Deng, Zhicheng Dou, and Ji-Rong Wen. 2023. Large language models for information retrieval: A survey. arXiv preprint arXiv:2308.07107.  
+Noah Ziems, Wenhao Yu, Zhihan Zhang, and Meng Jiang. 2023. Large language models are built-in autoregressive search engines. In ACL (Findings), volume ACL 2023 of Findings of ACL. Association for Computational Linguistics.
+
+# Appendix
+
+# A Prompt of LLM4CS
+
+Figure 4 shows a general illustration of the prompt of LLM4CS. The prompt consists of three parts, which are Instruction, Demonstration, and Input. The red part is for REW prompting, the blue part is for the RTR and RAR promptings, and the orange part is for RTR prompting. The green part is for our designed chain-of-thought.
+
+# B Case Study
+
+# B.1 Examples of Chain-of-Thought
+
+The example in Table 4 shows how our CoT takes effect. The CoT and Our Rewrite fields are generated by LLM4CS (REW + CoT). We can find that the generated CoT effectively illustrates the rationale behind the rewriting process. Please refer to our anonymously open-sourced repository for more examples.
+
+# B.2 Examples of Human Evaluation
+
+An example for each category is shown in Table 5. Please refer to our anonymously open-sourced repository for more examples.
+
+<table><tr><td>Instruction</td><td>For an information-seeking dialog, please help reformulate the question into rewrite that can fully express the user&#x27;s information needs without the need of context, but also generate an informative response to answer the question.</td></tr><tr><td>Demonstration</td><td>I will give you several example multi-turn dialogs, where each turn contains a question, a response, and a rewrite. The rewrite part begins with a sentence explaining the reason for the rewrite.Example #1:Question: What should I consider when buying a phone?Rewrite: This is the first turn. So the question should be rewritten as: What should I consider when buying a phone?Response: The design of the phone and the overall ...Question: Cool. Which one would you recommend?Rewrite: Based on Turn 1, you are inquiring about what should be considered when buying a phone. So the question should be rewritten as: Cool. Which smartphone would you recommend for me?Response: Just because a phone has everything...Example #2:...</td></tr><tr><td>Input</td><td>Your Task (only questions and responses are given):Context:Question: What was the basis of the Watergate scandal?Response: ...Question: ...Response: ...Current Question: So what happened to Nixon?Rewrite: So what happened to Nixon after the events of the Watergate scandal?(Now, you should give me the rewrite and an informative response of the **Current Question** based on the **Context**. The output format should always be: Rewrite:\$Reason. So the question should be rewritten as: \$Rewrite\nResponse: \$Response. Go ahead!)</td></tr><tr><td>Model Output</td><td>Rewrite: Based on all previous turns, Nixon was badly involved in the Watergate scandal. So the question should be rewritten as: So what happened to Nixon after the events of the Watergate scandal?Response: With the mounting evidence and loss...</td></tr></table>
+
+Figure 4: A general illustration of the prompt of LLM4CS. The prompt consists of three parts, i.e., Instruction, Demonstration, and Input. The red part is for REW prompting, the blue part is for the RTR and RAR prompts, and the orange part is for RTR prompting. The green part is for our designed chain-of-thought.
+
+<table><tr><td>CAsT-20 (Conversation #84)</td></tr><tr><td>Turn 1</td></tr><tr><td>Question: How does seed investment work for startups?Response: Seed money, sometimes known as seed funding or seed capital, is a form of securities offering in which an investor invests capital in exchange for an equity stake in ... Human Rewrite: How does seed investment work for startups?CoT: This is the first turn.Our Rewrite: How does seed investment work for startups?</td></tr><tr><td>Turn 2</td></tr><tr><td>Question: Tell me more about angel rounds. Response: Tell Me More was a National Public Radio interview show hosted by journalist Michel Martin ... offering sample segments, and soliciting user feedback. It began broadcasting on April 30, 2007. Human Rewrite: Tell me more about angel investment rounds. CoT: Based on the current question and the previous response, we are comparing angel rounds with crowdfunding.Our Rewrite: Tell me more about angel rounds in seed investment for startups.</td></tr><tr><td>Turn 3</td></tr><tr><td>Question: How does it compare with crowdfunding? Response: For crowdfunding of equity stock purchases, there is some research in social psychology that indicates that ... Managing communications with a large number of possibly disappointed investors and supporters can be a substantial, and potentially diverting, task. Human Rewrite: How does an angel investment round compare with crowdfunding? CoT: Based on the current question and the previous response, we are comparing angel rounds with crowdfunding.Our Rewrite: How does angel funding compare to crowdfunding for startups?</td></tr><tr><td>Turn 4</td></tr><tr><td>Question: How much is a typical A series round? Response: OpenGov has raised $47 million from venture capital and angel investors. A $3 million Series A round in 2012 included venture funds 8VC, Founder Collective, Valiant Capital, and a number of high-profile ... Human Rewrite: How much is a typical A series investment round? CoT: Based on the previous turns, we are talking about different types of funding for startups such as seed investment, angel rounds, and crowdfunding.Our Rewrite: How much funding is typically raised in an A series round for startups?</td></tr><tr><td>Turn 5</td></tr><tr><td>Question: What are the risks of investing? Response: Angel investors invest in early stage or start-up companies in exchange for an equity ownership interest. Angel ... Here is what angels particularly care about: The quality, passion, commitment, and integrity of the founders. Human Rewrite: What are the risks of investing in startups? CoT: Based on the previous turns, we are discussing different aspects of startup investments. Our Rewrite: What are the risks associated with investing in startups?</td></tr><tr><td>Turn 6</td></tr><tr><td>Question: What are the different types of instruments used by VCs? Response: ... Human Rewrite: What are the different types of instruments used by VCs? CoT: The user is asking about the different types of instruments used by venture capitalists. Our Rewrite: What are the various types of instruments employed by venture capitalists (VCs)?</td></tr></table>
+
+Table 4: An example to show how our CoT takes effect. The CoT and Our Rewrite fields are generated by LLM4CS (REW + CoT).
+
+# Good-H
+
+Question: How about replacing it instead?
+
+Human Rewrite: How much does it cost to replace a garage door opener?
+
+Our Rewrite: What is the cost of replacing a garage door opener?
+
+# Good-A
+
+Question: What's important for me to know about their safety?
+
+Human Rewrite: What's important for me to know about the safety of smart garage door openers?
+
+Our Rewrite: What are the safety features that I should consider when choosing a garage door opener?
+
+# Bad-O
+
+Question: Are there any famous foods?
+
+Human Rewrite: Are there any famous foods in Washington D.C.?
+
+Our Rewrite: Are there any famous foods?
+
+# Bad-C
+
+Question: What is its main economic activity?
+
+Human Rewrite: What is the main economic activity of Salt Lake City?
+
+Our Rewrite: What is the main economic activity in Utah?
+
+Table 5: Examples of the four categories in human evaluation.

2023/Large Language Models Know Your Contextual Search Intent_ A Prompting Framework for Conversational Search/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:badf0f61b50a4f7cac0dae8447d5d0161c907d0140666a9fbdb5c00a7494fccf
+size 939566

2023/Large Language Models Meet Harry Potter_ A Dataset for Aligning Dialogue Agents with Characters/e87156c1-b2b9-4f05-802a-2e45b7de9ff2_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf3ffbc5f4cfde31678afe5aa3c6015c2af869353996b36b14e7feafe7678904
+size 1019140

2023/Large Language Models Meet Harry Potter_ A Dataset for Aligning Dialogue Agents with Characters/full.md

@@ -0,0 +1,518 @@
+# Large Language Models Meet Harry Potter: A Bilingual Dataset for Aligning Dialogue Agents with Characters
+
+Nuo Chen†, Yan Wang‡, Haiyun Jiang‡, Deng Cai‡  
+Yuhan Li†, Ziyang Chen‡, Longyue Wang‡ and Jia Li†§  
+‡Tencent AI Lab
+
+†DSA, Hong Kong University of Science and Technology (Guangzhou), Hong Kong University of Science and Technology
+
+†chennuo26@gmail.com,§ {yanwang.branden@gmail.com, jialee@ust.hk}
+
+HPD Benchmark: https://nuochenpku.github.io/HPD.github.io
+
+# Abstract
+
+In recent years, Dialogue-style Large Language Models (LLMs) such as ChatGPT and GPT4 have demonstrated immense potential in constructing open-domain dialogue agents. However, aligning these agents with specific characters or individuals remains a considerable challenge due to the complexities of character representation and the lack of comprehensive annotations. In this paper, we introduce the Harry Potter Dialogue (HPD) dataset, designed to advance the study of dialogue agents and character alignment. The dataset encompasses all dialogue sessions (in both English and Chinese) from the Harry Potter series and is annotated with vital background information, including dialogue scenes, speakers, character relationships, and attributes. These extensive annotations may empower LLMs to unlock character-driven dialogue capabilities. Furthermore, it can serve as a universal benchmark for evaluating how well can a LLM aligning with a specific character. We benchmark LLMs on HPD using both fine-tuning and in-context learning settings. Evaluation results reveal that although there is substantial room for improvement in generating high-quality, character-aligned responses, the proposed dataset is valuable in guiding models toward responses that better align with the character of Harry Potter.
+
+# 1 Introduction
+
+With the emergence of dialogue-centric Large Language Models (LLMs) such as ChatGPT and GPT-4 (OpenAI, 2023), there has been a growing interest among researchers in exploiting the capabilities of these models to develop open-domain dialogue agents. A particularly exciting and challenging aspect of this pursuit involves aligning the behavior of these agents with a distinct character or individual (Kirk et al., 2023; Salemi et al., 2023).
+
+In this paper, using the Harry Potter series fiction as a prime example, we take a significant leap
+
+<table><tr><td>Dataset</td><td>Sce.</td><td>Att.</td><td>Re.</td><td>Dy.</td><td>Sl.</td></tr><tr><td>PchatbotW (2021)</td><td>×</td><td>√</td><td>×</td><td>×</td><td>×</td></tr><tr><td>PeDialog (2019)</td><td>×</td><td>√</td><td>×</td><td>×</td><td>×</td></tr><tr><td>KvPI (2020)</td><td>×</td><td>√</td><td>×</td><td>×</td><td>×</td></tr><tr><td>P-CHAT (2018)</td><td>×</td><td>√</td><td>×</td><td>×</td><td>×</td></tr><tr><td>WOW (2019)</td><td>√</td><td>√</td><td>×</td><td>×</td><td>×</td></tr><tr><td>Fri-QA (2019)</td><td>×</td><td>√</td><td>√</td><td>×</td><td>×</td></tr><tr><td>Focus (2021)</td><td>√</td><td>√</td><td>√</td><td>×</td><td>×</td></tr><tr><td>UltraChat(2023)</td><td>√</td><td>√</td><td>×</td><td>×</td><td>×</td></tr><tr><td>LaMP(2023)</td><td>√</td><td>√</td><td>×</td><td>×</td><td>×</td></tr><tr><td>Ours</td><td>√</td><td>√</td><td>√</td><td>√</td><td>√</td></tr></table>
+
+Table 1: Datasets Comparison. Sce., Att., Re., Dy. and Sl. denote Scenes, Attributes, Relations, Dynamic and Storyline, separately.
+
+towards aligning dialogue agents with characters in a story. Throughout this process, we discover that despite the assistance of super-powerful LLMs such as ChatGPT and GPT4, we encounter numerous challenges: (1) Firstly, the knowledge of LLMs is primarily dominated by real-world knowledge, which may diverge or even contradict to the character's story setting. In fact, ChatGPT and GPT4 have developed a substantial understanding of the Harry Potter series and the wizarding world behind it, yet they still occasionally produce hallucinations that contradict the story setting. For instance, GPT4-act Harry might attempt to visit Ron's house using the Hogwarts Express, despite the fact that this train exclusively operates between King's Cross Station and Hogwarts. (2) Secondly, LLMs encounter difficulties in interpreting intricate character relationships, particularly when multiple relationships coexist (e.g., friendships, romantic partnerships and competitors). (3) Finally, the most significant challenge stems from the LLMs' inability to accurately represent the impact of time on characters, leading to potential inconsistencies in their portrayal and development within the story.
+
+The dialogue on the right side of Figure 1 effectively illustrates the difficulties (2) and (3). We all know that Harry and Ron are very close friends, but
+
+![](images/6306f6e61aa619c1e40e4ae19cf86f62fe4956cc5d0f259d8b84232cad8552e5.jpg)  
+Figure 1: A conversation between Harry and Hermione selected from Book 4-Chapter 19 in the Harry Potter Series. In this example, we present the scene, timing of the conversation, relations and attributes of speakers. Texts in red refer to the expected response.
+
+they sometimes experience conflicts as well, like most adolescents. In Book 4-Chapter 19, Harry and Ron have a heated disagreement over the "Triwizard Tournament" registration (Ron believes that Harry has hidden the truth from him, but in reality, Harry has not), leading to a period of mutual avoidance. At this point, when Hermione suggests to Harry that they should find Ron for a Butterbeer, Harry's normal response would be reluctance to see Ron. However, even after providing an extensive context as a prompt, both ChatGPT and GPT-4 still generate responses that portray Harry and Ron as good friends, which is clearly inconsistent with the situation at that time.
+
+The appearance of the previously mentioned challenges is unsurprising, considering that current dialogue datasets intrinsically lack accurate annotations and guidelines for handling dynamic character relationships and attributes. As shown in Table 1, while some existing datasets (Zhang et al., 2018; Zheng et al., 2019; Jang et al., 2021; Ding et al., 2023) do include basic information about scenes, relationships, and attributes, the issue is that this information remains static and does not change over time. For example, if John and Harry are labeled as "friends", their relationship will remain as friends across all dialogue sessions.
+
+In this paper, we propose Harry Potter Dialogue (HPD), a dataset for facilitating the study of Dialogue Agents - Character aligning. This dataset encompasses all dialogue sessions from Harry Pot
+
+ter novels (English and Chinese versions)  ${}^{1}$  . In total,we obtain 1042 dialogue sessions for training (containing 1 positive response only) and 149 sessions for testing (containing 1-3 positive responses and 9 negative responses in average). We also annotate each conversation with essential background information that we believe is useful for aligning dialogue agents with Harry, including dialogue scenes, speakers, character relationships, and attributes. With the goal of giving a full picture of the speakers in dialogue, we have labeled each speaker with 12 types of relationships with Harry Potter and 13 types of attributes. Please note that even if the speakers in two different dialogues are identical, their relationships and attributes may change due to the story's progression.
+
+The main contributions of this paper can be summarized into three aspects:
+
+- We introduce the HPD dataset, designed to facilitate the study of aligning dialogue agents with characters. For each dialogue session in the Harry Potter novels, we provide all the background information that we believe may be helpful in assisting models to generate Harry Potter-aligned responses.  
+- This is a bilingual dialogue dataset that does not rely on machine translation. The only difference between languages is that they are
+
+based on Harry Potter novels in different languages. As a result, the data quality in both languages is comparable, making it suitable for investigating the impact of language differences on the task of aligning dialogue agents with characters.
+
+- The experimental results show that HPD can help LLMs, such as ChatGPT, better align with the behaviors of Harry Potter. However, the degree of alignment is still far from the level of human experts, indicating ample room for further improvement.
+
+# 2 Task Definition
+
+We use the Harry Potter novels as our test-bed with the aim of aligning dialogue agents with Harry Potter in a story. The generated responses of such a dialogue agent should be not only relevant to the context, but also seem like something Harry would say at the time and scene.
+
+Figure 1 shows some main factors that affect behaviors of Harry in a conversation. The first factor is the conversation history, which is the most important factor that determines Harry's response. The scene, which is the second factor, provides details about the motivation (Hermione invites Harry to have a butterbeer with Ron in the Three Brromsticks) of this dialogue. The third factor is the participants' information (attributes and relations), obviously, Harry will say very different things to different characters, such as Malfoy and Hermione. The latter two factors belong to the background information and are dynamically determined by the timing of this dialogue (Book 4-Chapter 19), and they are continuously varied over the storyline.
+
+Formally, the task of aligning dialogue agents with characters in a story can be defined as follows: Given a dialogue history  $\mathbf{H}$ , corresponding dialogue scene  $\mathbf{S}$  and participants information  $\mathbf{P}$  as input, which evolve depending on the development of storyline. The dialogue agent is supposed to generate a response  $\mathbf{Y} = \{y_1, y_2, \dots, y_n\}$ :
+
+$$
+\mathcal {Y} = \underset {Y} {\operatorname {a r g m a x}} P (\mathbf {Y} | \mathbf {H}, \mathbf {S}, \mathbf {P})
+$$
+
+$\mathbf{Y}$  is supposed to be not only fluent and natural, but also highly relevant to  $\mathbf{S}$  and  $\mathbf{P}$ .
+
+# 3 Dataset Construction
+
+A high-quality dataset including all pertinent information in Section 2 is the prerequisite for aligning
+
+dialogue agents with characters in a story. Unfortunately, so far there are currently no publicly available datasets that provide information about the dialogue scene and participants. To facilitate the study of this task, we construct a new dataset from the popular fictions Harry Potter Series, in the hopes of creating a Harry Potter-aligned dialogue agent. All dialogue sessions that Harry participates in are collected in our dataset, along with fine-grained annotated dialogue scenes and participant information. We recruit four avid Harry Potter fans (professional annotators, no crowd-sourcing) for the annotation work in this study.
+
+We collect three parts of information to construct our dataset, as shown in Figure 1: 1) The dialogue part (Section 3.1) contains all utterances in the dialogue sessions, as well as the speaker's name of each utterance. 2) The scene part (Section 3.2) includes the summarization of the text around the dialogue session. 3) Finally, the speaker information part (Section 3.3), which consists of attributes and relations of characters, is shown in the left part of Figure 1. Please note that these scenes, attributes, and relations are time-sensitive, they may change as the storyline goes on, so we should annotate them session by session.
+
+# 3.1 Dialogue Construction
+
+Dialogue sessions in the books are divided into a training set and a test set. Their main difference is that each training session contains only one positive response, while each test dialogue session consists of at least 1 positive response (human-written ground-truth) and 9 negative responses. We hope this test set may support the evaluation of both dialogue generation and retrieval.
+
+Training Set We request the annotators to extract all multi-turn dialogues from the books. Besides, the speaker name of each utterance in the session is labeled as well.
+
+Test Set The effectiveness of dialogue models should be evaluated using a well-designed test set. However, if we directly select the test dialogue sessions from the books, it may meet serious knowledge leakage problems, i.e., the fact that we evaluate in the test set also presents in the pre-training corpus. To prevent this problem, we deliberately design a test set in the following steps:
+
+- First, we manually select some raw dialogues that meet the following requirements: (i) Dia
+
+![](images/37f5497380a4c1bc27d735d4a241241a25fe2b00e583ca88c3c395569009bcdd.jpg)  
+Figure 2: Data examples from two different timings: Book 1-Chapter 12 and Book 6-Chapter 7. Words in color denote the changed relations and attributes.
+
+logues with only one speaker, which contain only one or two sentences; (ii) Dialogues, in which there is no response from other speakers to the last question. For these samples, we pick out dialogues that are relevant to Harry and can be answered from Harry's perspective to construct the test set.
+
+- Second, since directly composing a high-quality Harry-like response is quite challenging for the annotators, we try to alleviate the difficulty with the help of LLMs. Specifically, we utilize ChatGPT and GPT-4 to generate potential responses for each selected dialogue session under in-context learning settings. Subsequently, each model is needed to predict 5 responses as the candidates at least.  
+- Third, we ask three annotators to select the most reasonable responses from the predictions as the positive response (ground-truth)<sup>2</sup> while the remaining samples are taken as negatives. A holdout annotator is responsible for integrating the annotations. If the selected response is not comprehensive enough or lacks certain information, rewriting is required. Furthermore, if all predictions are considered unreasonable, the holdout annotator would write a response from scratch.
+
+This setup has advantages over employing a single annotator to label predictions: One annotator may regard some predictions as positive responses while the other may label them as negatives. Hence, the holdout annotator can be used to guarantee the
+
+quality of annotated answers when disagreements appear. These strategies alleviate spam and bias, and thus get a high-agreement dataset. Finally, we manually double-check and revise mistakes to further ensure the data quality.
+
+# 3.2 Scenes Construction
+
+In order to offer accurate location information and textual details for each dialogue, we further annotate scenes. We assume that the texts that surround and immediately relate to the dialogue in the novel provide ample scene information. Therefore, we initially instruct the annotators to extract these texts from the novel. Subsequently, to leverage the capabilities of GPT-4, we employ it to summarize the extracted texts. To ensure the utmost precision and accuracy, two skilled annotators are required to meticulously calibrate the summarized texts, resulting in the final scene data.
+
+It is worth noting that we deliberately abstained from directly utilizing the extracted texts as scenes, primarily due to two compelling reasons: Firstly, the inherent length of the original text exceeds the maximum limit imposed by most LLMs, rendering it impractical for our purposes. Secondly, in real-world scenarios, the attainment of such exceptionally high-quality scenes is highly improbable, as scene information typically originates from video captions or rule-based templates.
+
+# 3.3 Attributes and Relations Construction
+
+One of the most important and appealing properties of our benchmark is the fine-grained annotated character information, which includes the attributes and relations of the characters. With the goal of
+
+providing in-depth and comprehensive character information, we collect 13 fine-grained attributes and 12 relations. We take into account all of these attributes and relationships that are dynamically evolving as the story goes on. Thus, we annotate them chapter by chapter. Considering not all characters are essential to understanding and driving the story in Harry Potter series, we target 113 important characters to annotate their attributes and relations, such as Harry, Ron.
+
+We divide the attributes into two categories: (1) inborn; (2) nurture. The former denotes some innate attributes or abilities, which contains  $\{Gender, Age, Lineage, Talents, and Looks\}$ . The latter refers to properties acquired through efforts, including  $\{Achievement, Title, Belongings, Export, Hobby, Character, Spells and Nickname\}$  (some cases are presented in Figure 2). In total, we collect 13 attributes for each character, which basically cover most properties in the Harry Potter series.
+
+The relations between Harry and other characters can be classified into binary relations and discrete relations. The former include 8 types, which are {Friend, Classmate, Teacher, Family, Lover, Opponent, Teammate, Enemy}. Multiple binary relations can exist between two characters. Harry and Ron, for instance, are friends, classmates, and teammates in the Quidditch team (In Book-6 only).
+
+In order to fully represent the relationship between two characters, we need to know not only the type of relationship (the binary relations) they have but also their familiarity and affection for each other. We annotate affection and familiarity as 4 types of discrete relations: (1) Harry's Familiarity with someone, (2) Harry's Affection for someone, (3) someone's Familiarity with Harry, and (4) someone's Affection for Harry.
+
+The difference between these types can be illustrated by the following two examples: 1) Draco Malfoy hates Harry, but he is also familiar with Harry. So his Affection for Harry is low but his Familiarity with Harry is high. 2) In addition, ever since Harry lost his parents, Dumbledore has shown great concern for Harry. Hence, his Familiarity and Affection for Harry are high, while Harry's Familiarity with Dumbledore is relatively low.
+
+Figure 2 is another example of how attributes and relationships have changed over the course of the story. In Book 1-Chapter 7, Harry has just met Ron on the train to Hogwarts and is an alien in the wizarding world. So his affection and familiarity
+
+![](images/4c512e6d80833fea770628ad40804778893afdc418e863b64ef3c74dc1998015.jpg)  
+Figure 3: Affection Definition and Examples.
+
+![](images/6315f2918b71fbca0ee41c45d301648b31e0bfae78a2d358f8160a75bfdd625d.jpg)  
+Figure 4: Familiarity Definition and Examples.
+
+with Ron are relatively low (1 and 2, respectively), and he isn't aware of any spells. As the story progresses, however, in Book 6-Chapter 7, he is a full-fledged wizard, and Ron is his best friend. So their affection and familiarity are high at this time (7 and 8, respectively). Harry also masters a lot of spells such as Expecto Patronum and Expelliarmus, and has some equipments such as his broomstick Firebolt and the invisibility cloak.
+
+Affection Definition Affection is rated on 21 levels, ranging from -10 to 10, where -10 and 10 indicate the lowest and highest affection, respectively. A positive affection level means the character has a positive relationship with Harry and vice versa.
+
+Figure 3 shows the detailed definition and some examples of different Affection levels. For example, 1 refers to "someone meets Harry for the first
+
+<table><tr><td>Statistics</td><td>Train</td><td>Test</td></tr><tr><td colspan="3">per dialogue</td></tr><tr><td>Average Turns</td><td>13.8</td><td>7.8</td></tr><tr><td>Maximum Speakers</td><td>20</td><td>8</td></tr><tr><td>Minimum Speakers</td><td>2</td><td>2</td></tr><tr><td colspan="3">per utterance</td></tr><tr><td>Average Length</td><td>32.9</td><td>28.3</td></tr><tr><td>Maximum Length</td><td>77</td><td>26</td></tr><tr><td>Minimum Length</td><td>3</td><td>3</td></tr><tr><td>Total Dialogues</td><td>1042</td><td>149</td></tr></table>
+
+Table 2: Data statistics of collected dialogues.
+
+time", when Harry first met Ron and Hermione in Book 1, Harry's Affection to both is 1. As for  $-10$ , it means the deepest hatred, and the most obvious example is Harry and Voldemort.
+
+Familiarity Definition Similarly, we also rate Familiarity with a 11-point scale, which ranges from 0 to 10, where 10 is the highest and 0 is the lowest. We present detailed definitions for each Familiarity level in Figure 4. Concretely, 0 denotes strangers, and 10 denotes very close friends who stay together for many years and are very familiar with each other's habits, secrets, and temperaments. Ron in Book 7 meets this condition.
+
+Annotation During annotation, we ask annotators to annotate Attributes and Relations following the above definitions and examples. To eliminate the effects of individual bias, we ask three of them to repeatedly label the Attributes and Relations chapter by chapter. And a holdout annotator (senior) is responsible for the quality of annotated data when the former three annotators have disagreements. Finally, we manually re-check all annotated data and revise some controversial annotations to further control the data quality.
+
+Claim Notice that we hope to provide as rich character information as possible for the community, even if some of them seem redundant in this work. Therefore, we collect 13 types of attributes for each character and 12 types of relations in the collected HPD. We leave plenty of opportunity for other research communities to investigate which information is helpful in their work. For example, it can support other tasks such as the sentiment analysis of Harry Potter. One may not need to include all the fine-grained annotated information in his/her study, because it is still challenging.
+
+# 3.4 Data Statistics
+
+The detailed statistics of dialogues are shown in Table 2. The training set and test set contain 1042 and 149 dialogue sessions, respectively. Of note, we initially collect 1471 dialogues for constructing the training set, and we filter out those dialogues that are without Harry, leading to 1042 conversations for consideration. Most of the conversations involve multiple speakers, with the maximum one including up to 20 speakers. It is obviously a serious challenge to the dialogue agents.
+
+# 4 Experiments
+
+In this section, we conduct extensive experiments to investigate 1) whether Annotations in HPD can assist LLMs in aligning with characters, and 2) if so, whether the degree of alignment can reach human-level performance. We will discuss these questions in section 4.3.
+
+# 4.1 Baselines
+
+We build multiple strong baselines in our experiments, which can be divided into two types: generation-based and retrieval-based systems. Due to the page limit, here we only introduce the generation task in HPD. For details of retrieval tasks in HPD, please refer to Appendix B and D.
+
+Models We implement different generative models in two ways: fine-tuning and in-context learning. For the former, we fine-tune  $\mathbf{Alpaca}^3$  (6B) and ChatGLM-6B on our dataset. For the latter, we deliberately design prompts for GPT3 (Brown et al., 2020) (text-davinci-002), ChatGPT (gpt3.5-turbo), and ChatGLM (Zeng et al., 2022) (chatglm-130B). Furthermore, to explore the effect of annotated fine-grained background knowledge, some methods (Alpaca, GPT3, and ChatGPT) are implemented in two different settings: 1) base setting with a prompt that only includes task description, one dialogue example, and dialogue history; 2) rich-persona setting (denoted as Per-Model) with a prompt contains all annotated background information in HPD as in-context learning exemplars. The detailed prompts can be found in Appendix C (Table 8 for base setting, and Table 10 for rich-persona setting).
+
+<table><tr><td rowspan="2">Model</td><td colspan="3">Automatic Evaluation</td><td colspan="3">GPT-4 Evaluation</td><td colspan="3">Human Evaluation</td></tr><tr><td>Bleu-1</td><td>Dist.1</td><td>Rough-L</td><td>Relv.Sce.</td><td>Relv.Attr.</td><td>Relv.Re.</td><td>Relv.Sce.</td><td>Relv.Attr.</td><td>Relv.Re.</td></tr><tr><td colspan="10">Fine-tuning</td></tr><tr><td>CGLM*-6B</td><td>1.2</td><td>18.4</td><td>3.4</td><td>1.21</td><td>1.10</td><td>1.01</td><td>1.34</td><td>0.67</td><td>0.67</td></tr><tr><td>Alpaca</td><td>2.0</td><td>30.2</td><td>10.8</td><td>3.37</td><td>1.19</td><td>1.51</td><td>2.03</td><td>1.19</td><td>1.51</td></tr><tr><td>Per-Alpaca</td><td>12.2</td><td>20.4</td><td>13.8</td><td>3.51</td><td>1.34</td><td>1.51</td><td>4.04</td><td>1.34</td><td>1.51</td></tr><tr><td colspan="10">In-context learning</td></tr><tr><td>CGLM*-130B</td><td>14.2</td><td>25.1</td><td>15.7</td><td>6.71</td><td>6.01</td><td>6.04</td><td>7.38</td><td>6.04</td><td>5.37</td></tr><tr><td>GPT3</td><td>9.8</td><td>23.0</td><td>14.6</td><td>2.68</td><td>2.68</td><td>3.36</td><td>4.70</td><td>4.70</td><td>4.03</td></tr><tr><td>Per-GPT3</td><td>22.6</td><td>20.1</td><td>16.5</td><td>12.75</td><td>6.04</td><td>3.36</td><td>12.08</td><td>6.71</td><td>5.37</td></tr><tr><td>ChatGPT</td><td>33.1</td><td>19.0</td><td>20.2</td><td>33.56</td><td>38.92</td><td>37.58</td><td>32.89</td><td>34.23</td><td>32.89</td></tr><tr><td>Per-ChatGPT</td><td>33.6</td><td>19.4</td><td>22.4</td><td>37.58</td><td>42.95</td><td>46.98</td><td>35.57</td><td>45.63</td><td>48.99</td></tr></table>
+
+Table 3: A comprehensive evaluation of all baselines on HPD test set. Per-Model means the model with prompts in rich-persona setting. Here we report the percentage of generated responses ranked as the best one for each dialogue session (top-1 ranking) in GPT-4 and human evaluation. CGLM refers to ChatGLM. We report the average ranks of these LLMs in Figure 5. Dist.1 is short for Distinct-1.
+
+# 4.2 Evaluation Metric
+
+To comprehensively measure whether the texts generated by various models exhibit a Harry Potter-aligned tone, we employ three different kinds of evaluation methods: reference-based (automatic), GPT-4 based, and human-based. Although some existing studies (Liu et al., 2023; Peng et al., 2023) claims that GPT-4 is good-enough to assess the quality of generated texts, we found that human judgment is still the most thorough and realistic assessment of whether the generated text is Harry Potter-aligned in our task.
+
+Reference-based Metrics We employ Rough-L (Lin, 2004), Bleu-1 (Papineni et al., 2002), Distinct-1 as our metrics to validate the relevance and diversity of the generated responses.
+
+GPT-4 Evaluation We evaluate the persona's consistency based on the three criteria: Relevance with the Scene (Relv.Sce.), Relevance with the Attributes (Relv.Attn.) and Relevance with the Relations (Relv.Re.). Considering the potential drawbacks of having GPT-4 to assign scores for all responses individually, such as the risk of confusing GPT-4's comprehension, we have opted for an alternative approach. We instruct GPT-4 to rank the generated texts based on above distinct criteria. This ranking process allows us to evaluate the performance of different models effectively. The detailed prompts are in Appendix C, Table 9.
+
+Human Evaluation In light of the discovery of some poor GPT-4 annotation cases, we further instructed our annotators to revise and rectify the ranking results of GPT-4 for each test data, leading
+
+to human-evaluation results. During the annotation, annotators were allowed to consult the original text for reference when ranking. Moreover, as long as at least one annotator among the three made adjustments to GPT-4's ranking results, we would adopt the modified results as the human evaluation results. If multiple annotators made revisions to a single result, we would take the average of their modifications as the final human evaluation result.
+
+# 4.3 Results
+
+Whether the HPD can assist LLMs in aligning with characters? To answer this question, we present the results of these baselines in Table 3, and observe that across all GPT-4 and human-based metrics, methods perform significantly better in the rich-persona setting compared to their performance in the base setting. This improvement, obviously, should be attributed to the additional background information provided by the HPD.
+
+Additionally, for ChatGPT, the improvement in persona consistency-related metrics, including Relv.Att. and Relv.Re., is more pronounced (for example, compared to ChatGPT, Per-ChatGPT has increased by  $11.4\%$  and  $16.1\%$  in these two metrics, respectively). This is mainly because, in most cases, the dialogue history has some overlap with the scene, and ChatGPT can be consistent with the scene by merely using the dialogue history. For GPT-3 and Alpaca, their understanding of dialogue history is not as strong as that of ChatGPT, so their improvements in Relv.Sce. are still noticeable.
+
+Models VS. Human Experts Intuitively, we may further raise a question: How good are the gener
+
+<table><tr><td>Category</td><td>Relv.Sce.</td><td>Relv.Attr.</td><td>Relv.Re.</td><td>Ave.</td></tr><tr><td>WIN</td><td>77.85</td><td>77.18</td><td>73.15</td><td>76.06</td></tr><tr><td>TIE</td><td>1.34</td><td>4.03</td><td>8.05</td><td>4.47</td></tr><tr><td>LOSE</td><td>20.81</td><td>18.79</td><td>18.79</td><td>19.46</td></tr></table>
+
+Table 4: Comparison between Per-ChatGPT and Human experts. The win/tie/lose proportions of human-written responses on different metrics are reported.  
+
+<table><tr><td>Scene: Harry and his family visit the reptile house at the zoo, where Harry strikes up a conversation with a boa constrictor from Brazil... Harry is blamed for talking to the snake and causing the incident.</td></tr><tr><td>Dialogue Position: Book1-chapter2</td></tr><tr><td>Speakers: Dudley, Harry...Piers</td></tr><tr><td>Dialogue: Dudley: Make it move...</td></tr><tr><td>the snake: I get that all the time.</td></tr><tr><td>Harry: I know, Where do you come from, anyway?</td></tr><tr><td>the snake: Boa Constrictor, Brazil.</td></tr><tr><td>Harry: Was it nice there? Oh, I see — so you’ve never been to Brazil?</td></tr><tr><td>keeper of the reptile house: DUDLEY! MR. DURSLEY!</td></tr><tr><td>COME AND LOOK AT THIS SNAKE! YOU WON’TE BELIEVE WHAT IT’S DOING!</td></tr><tr><td>the snake: Brazil, here I come. . . . Thanksss, amigo.</td></tr><tr><td>Piers: Harry was talking to it, weren’t you, Harry?</td></tr></table>
+
+<table><tr><td>GPT-3: I was just asking it where it came from. It&#x27;s not like I was trying to make friends with it or anything.</td></tr></table>
+
+<table><tr><td>Per-GPT3: I didn’t mean to cause any trouble, I was just curious about where the snake came from. I didn’t know it would escape.</td></tr></table>
+
+<table><tr><td>ChatGPT: Yes, I was just curious about where it came from and if it had ever been to Brazil. But Dudley, you shouldn&#x27;t be so rude to the keeper of the reptile house. He was just trying to show us something amazing.</td></tr></table>
+
+<table><tr><td>Per-ChatGPT: Yes, I was. But I didn&#x27;t make the glass disappear or tell the snake to escape. It was just a harmless conversation. I didn&#x27;t mean for any of this to happen.</td></tr></table>
+
+<table><tr><td>Annotated Response: No, I wasn&#x27;t talking to it. I was just looking at it and reading the sign about where it came from. You two must have imagined things.</td></tr></table>
+
+Table 5: Examples of generated responses by baselines. Here, we have omitted the attributes and relationship information between the characters.
+
+ated texts? Are they as good as the ground-truth? To explore this, we further recruit another three annotators to compare the Per-ChatGPT-generated responses with the human-written responses in the test set. The experimental results, as shown in Table 4, are surprising. Contrary to our common sense that ChatGPT has already reached human-level performance in conversations, humans show a clear advantage in the task of aligning with Harry Potter. In all metrics, the proportion of PreChatGPT outperforming or being on par with humans even lower than  $30\%$ . As we analyzed in Section 1, the existing LLMs are still far from being able to align with specific characters.
+
+# 4.4 Case Study
+
+As the case in Table 5, the scene of selected dialogue is about "Harry finds he is able to converse with snakes." and Harry is facing questioning from Pears, who often bullies Harry with Dudley. He dislikes and even feels somewhat afraid of the Dudley family and Pears. At that time, Harry doesn't know anything about the wizarding world, and hasn't yet known about or gone to Hogwarts. Hence, Harry is likely trying to downplay the situation, keep his secret (he can talk with snakes) and avoid potential conflict with Piers and Dudley. However, all models cannot capture such a complicated relationship and fail to generate real Harry Potter-aligned responses. Most models generate responses that start with "Yes, I was..." or "I was", which are dull and clearly contradict Harry's actual intention.
+
+Additional examples can be found in Appendix E, Table 11. They demonstrate that when dialogues necessitate a deep understanding of the current context or involve a sudden shift in Harry's relationships with other characters, current LLMs' behavior significantly deviates from Harry's actual actions. This is, obviously, far from satisfactory. Although the Harry Potter series is already included in the training corpus of most LLMs (they know the basic information about characters and their world), relying solely on 'next word prediction' may be insufficient to fully understand the nuanced knowledge embedded within the story behind Harry Potter.
+
+# 5 Conclusion
+
+In this paper, we propose a new benchmark named Harry Potter Dataset (HPD) to promote aligning dialogue agents with characters in a story. Unlike existing datasets, HPD not only contains interesting dialogues, but also scenes, character attributes and relations that are dynamically changed as the storyline goes on. It also provides a well-designed test set to facilitate the evaluation of both generation-based and retrieval-based dialogue agents. Results and case studies show that powerful LLMs are still far from human expectations, proving there is ample room for improvement. Generally, HPD offers several open research problems in character aligning, such as how to build automatic evaluation metrics for personalized text generation, exploring effectively prompting LLMs for character aligning in a story. We hope HPD can play an crucial role in moving through cracking them.
+
+# Ethical Statement
+
+To avoid the potential issue of using Harry Potter novels, we promise the annotated dataset is developed for non-commercial use. Moreover, we only provide the line number and page number of each collected dialogue in Harry Potter rather than the detailed content of each dialogue session. We further supply the script to extract corresponding raw dialogue data from the novels according to the provided line and page numbers, in which the data format is the same as the data examples in Table 11. As for the annotated character attributes and relations, we have our own copyright and will release for research communities.
+
+# Limitations
+
+The main target of this paper is towards building dialogue agents for characters in a story. In this paper, we present a new benchmark named Harry Potter Dialogue (HPD) in the hope of creating a Harry Potter-aligned dialogue agent. The significant feature of HPD is that it contains detailed scenes and fine-grained attributes and relations of each speaker which are dynamically changed as the storyline goes on. More generally, we expect the core idea of this paper can give insights into other research communities that want to build effective person-like chatbots in the virtual world. Our well-designed test sets could even support RLHF training of characterized dialogue agents. Our fine-grained annotated knowledge also can be used to build other tasks such as sentiment analysis and reading comprehension of Harry Potter. Admittedly, the data in the proposed dataset from the Harry Potter Series is restricted to a specific area, that is, Harry Potter Magic World. Considering the high cost of annotation, our character relation annotation work is restricted to Harry Potter. These concerns warrant further research and consideration when utilizing this work to build intelligent person-like dialogue systems in the virtual world.
+
+# References
+
+Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess,
+
+Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. CoRR, abs/2005.14165.  
+Nuo Chen, Hongguang Li, Yinan Bao, Junqing He, Xinshi Lin, Qi Yang, Jianfeng Liu, Ruyi Gan, Ji-axing Zhang, Baoyuan Wang, et al. 2023. Orca: A few-shot benchmark for chinese conversational machine reading comprehension. arXiv preprint arXiv:2302.13619.  
+Nuo Chen, Chenyu You, and Yuexian Zou. 2021. Self-supervised dialogue learning for spoken conversational question answering. In *Interspeech*, pages 231-235. ISCA.  
+Cristian Danescu-Niculescu-Mizil and Lillian Lee. 2011. Chameleons in imagined conversations: A new approach to understanding coordination of linguistic style in dialogs. In CMCL@ACL, pages 76-87. Association for Computational Linguistics.  
+Emily Dinan, Stephen Roller, Kurt Shuster, Angela Fan, Michael Auli, and Jason Weston. 2019. Wizard of wikipedia: Knowledge-powered conversational agents. In ICLR (Poster). OpenReview.net.  
+Ning Ding, Yulin Chen, Bokai Xu, Yujia Qin, Zhi Zheng, Shengding Hu, Zhiyuan Liu, Maosong Sun, and Bowen Zhou. 2023. Enhancing chat language models by scaling high-quality instructional conversations. arXiv preprint arXiv:2305.14233.  
+Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685.  
+Yoonna Jang, Jungwoo Lim, Yuna Hur, Dongsuk Oh, Suhyune Son, Yeonsoo Lee, Dong-Hoon Shin, Seungryong Kim, and Heuseok Lim. 2021. Call for customized conversation: Customized conversation grounding persona and knowledge. CoRR, abs/2112.08619.  
+Hannah Rose Kirk, Bertie Vidgen, Paul Röttger, and Scott A Hale. 2023. Personalisation within bounds: A risk taxonomy and policy framework for the alignment of large language models with personalised feedback. arXiv preprint arXiv:2303.05453.  
+Satwik Kottur, Xiaoyu Wang, and Vítor Carvalho. 2017. Exploring personalized neural conversational models. In *IJCAI*, pages 3728-3734. ijcai.org.  
+Jiwei Li, Michel Galley, Chris Brockett, Georgios P. Spithourakis, Jianfeng Gao, and William B. Dolan. 2016. A persona-based neural conversation model. In ACL (I). The Association for Computer Linguistics.  
+Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain. Association for Computational Linguistics.
+
+Yongkang Liu, Shi Feng, Daling Wang, Yifei Zhang, and Hinrich Schütze. 2023. Evaluate what you can't evaluate: Unassessable generated responses quality. arXiv preprint arXiv:2305.14658.  
+OpenAI. 2023. Gpt-4 technical report.  
+Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, July 6-12, 2002, Philadelphia, PA, USA, pages 311-318. ACL.  
+Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. 2023. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277.  
+Hongjin Qian, Xiaohe Li, Hanxun Zhong, Yu Guo, Yueyuan Ma, Yutao Zhu, Zhanliang Liu, Zhicheng Dou, and Ji-Rong Wen. 2021. Pchatbot: A large-scale dataset for personalized chatbot. In SIGIR, pages 2470-2477. ACM.  
+Alireza Salemi, Sheshera Mysore, Michael Bendersky, and Hamed Zamani. 2023. Lamp: When large language models meet personalization. arXiv preprint arXiv:2304.11406.  
+Haoyu Song, Yan Wang, Weinan Zhang, Zhengyu Zhao, Ting Liu, and Xiaojiang Liu. 2020. Profile consistency identification for open-domain dialogue agents. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6651-6662.  
+Tengtao Song, Nuo Chen, Ji Jiang, Zhihong Zhu, and Yuexian Zou. 2023. Improving retrieval-based dialogue system via syntax-informed attention. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE.  
+Thomas Wolf, Victor Sanh, Julien Chaumond, and Clement Delangue. 2019. Transfertransfo: A transfer learning approach for neural network based conversational agents. CoRR, abs/1901.08149.  
+Chen Xu, Piji Li, Wei Wang, Haoran Yang, Siyun Wang, and Chuangbai Xiao. 2022. Cosplay: Concept set guided personalized dialogue generation across both party personas. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 201-211.  
+Zhengzhe Yang and Jinho D. Choi. 2019. *Friendsqa: Open-domain question answering on TV show transcripts*. In SIGdial, pages 188-197. Association for Computational Linguistics.  
+Chenyu You, Nuo Chen, Fenglin Liu, Dongchao Yang, and Yuexian Zou. 2020. Towards data distillation for end-to-end spoken conversational question answering. CoRR, abs/2010.08923.
+
+Chenyu You, Nuo Chen, and Yuexian Zou. 2021. Self-supervised contrastive cross-modality representation learning for spoken question answering. In EMNLP (Findings), pages 28-39. Association for Computational Linguistics.  
+Aohan Zeng, Xiao Liu, Zhengxiao Du, Zihan Wang, Hanyu Lai, Ming Ding, Zhuoyi Yang, Yifan Xu, Wendi Zheng, Xiao Xia, et al. 2022. Glm-130b: An open bilingual pre-trained model. arXiv preprint arXiv:2210.02414.  
+Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and Jason Weston. 2018. Personalizing dialogue agents: I have a dog, do you have pets too? In ACL (1), pages 2204-2213. Association for Computational Linguistics.  
+Wei-Nan Zhang, Qingfu Zhu, Yifa Wang, Yanyan Zhao, and Ting Liu. 2019. Neural personalized response generation as domain adaptation. World Wide Web, 22(4):1427-1446.  
+Yinhe Zheng, Guanyi Chen, Minlie Huang, Song Liu, and Xuan Zhu. 2019. Personalized dialogue generation with diversified traits. CoRR, abs/1901.09672.  
+Yinhe Zheng, Rongsheng Zhang, Minlie Huang, and Xiaoxi Mao. 2020. A pre-training based personalized dialogue generation model with persona-sparse data. In AAAI, pages 9693-9700. AAAI Press.
+
+# A Related Work
+
+Recently, building personalized dialogue systems draw a lot of attention from research communities. Aiming for promoting this area, several efforts and benchmarks (Chen et al., 2021; Danescu-Niculescu-Mizil and Lee, 2011; Zheng et al., 2019; Yang and Choi, 2019; Song et al., 2020; Zheng et al., 2020; You et al., 2020; Chen et al., 2023; You et al., 2021; Xu et al., 2022) have been made, demonstrating promising results for endowing personal style into dialogue systems. Some initial efforts (Danescu-Niculescu-Mizil and Lee, 2011) aimed at modeling characters from movies.
+
+Further developments provide personas via two types: implicit and explicit personalization. In the former streams (Kottur et al., 2017; Li et al., 2016; Zhang et al., 2019), each speaker's personality information can be compressed as the persona embeddings. In this manner, the existing issue of these methods makes it hard to explain their effectiveness. For the latter (Wolf et al., 2019; Song et al., 2020), the personal information are provided as: (1) dense personas, such as speaker profile or text-described personas; (2) sparse personas, including some personality traits. For example, personas from (Zheng et al., 2019) are formulated as key-value pairs: "Age:xx, Gender:xx, Location: xx".
+
+More recently, several efforts (Dinan et al., 2019; Yang and Choi, 2019; Jang et al., 2021; Song et al., 2023) incorporated scenes and relations knowledge into each dialogue session for encouraging more real personalized conversation. Yang and Choi (2019) presented a open-domain question answering dataset excerpted from *Friends Series*, where each dialogue involves multiple speakers and their relations. Jang et al. (2021) proposed FoCUS dataset where the customized responses are generated based on the user's persona and Wikipedia background knowledge.
+
+In this paper, our goal is to align dialogue agents with characters in a story, which requires modeling scenes and speaker information that are dynamically changed as the storyline goes on. However, the personality settings of current studies are static, and are not changed with scenes or times changing. Therefore, we present HPD: Harry Potter Dialogue Dataset, aiming for creating Harry Potter-alignd dialogue agent. In detail, we annotate detailed scenes, attributes and relations of each speaker over given dialogues to help the model have a deeper under
+
+standing of the dialogue background information.
+
+# B Baselines Setup
+
+In the following, we briefly introduce each model and describe the training and test details.
+
+# B.1 Baselines
+
+BERT-FP is a commonly-used strong retrieval-based dialogue system, which devises several posttraining objectives. When fine-tuning BERT-FP, given  $n - 1$  utterances in each dialogue session, the model is required to find the ground-truth response from candidate answers. Concretely, we first posttrain BERT-FP in Harry Potter novels and then fine-tune the resulting model in the collected HPD.
+
+# B.2 Experimental Setup
+
+Our experimental settings can be see in Table 6. Notice that, considering the memory cost, we utilize LoRA-tuning (Hu et al., 2021) and prompts don't contain any examples.
+
+# C Prompts
+
+In this section, we present three different prompts in our experiments: Table 8 shows the prompts of the base setting which only include task description, a dialogue example and the input dialogue. Table 9 presents the ranking prompts for GPT-4 evaluation. Table 10 illustrates the rich-persona setting prompts.
+
+# D Retrieval Task
+
+Automatic Metrics For evaluating the retrieval-based model, we also employ some common metrics: MAP (mean average precision), MRR (mean reciprocal rank), and  $\mathrm{P@1}$  (precision at one). Recall also be considered, which is used as R10@k, which implies that the correct response exists among the top k candidates out of the ten candidate responses.
+
+Automatic Evaluation BERT-FP also performs poorly on  $P@1$  score (25.9%) and MAP score (46.8%) in the retrieval-based task. The results show the current state-of-the-art retrieval-based model also can not handle the challenge of our benchmark, and thus, there is ample room for future improvement.
+
+<table><tr><td>Parameter</td><td>Alpaca</td><td>Per-Alpaca</td><td>ChatGLM-6B</td><td>BERT-FP</td></tr><tr><td>Batch size</td><td>4</td><td>4</td><td>1</td><td>32</td></tr><tr><td>Learning Rate</td><td>1e-4</td><td>1e-4</td><td>3e-5</td><td>8e-5</td></tr><tr><td>Epoch</td><td>5</td><td>5</td><td>5</td><td>20</td></tr><tr><td>Max Length</td><td>512</td><td>512</td><td>650</td><td>256</td></tr></table>
+
+Table 6: Hyper-parameters setup in fine-tuning.
+
+![](images/6db0191cf635e6ee47c02ad572a417d6fec51e0f4a1b1e2cd2421181dd784597.jpg)
+
+![](images/71f2bfca24f421f7db5caba40ba11a1bbde36113745c9bcc15c68fe8497f85d1.jpg)  
+(a) GPT-4 Evaluation
+
+![](images/e8a4ea0a74d1ecbd84a2fd2e5f08fc72ac0a691dbc5b0708912b21573fcaab00.jpg)
+
+![](images/491c6282b7333db83d08b305573d4105ecdaed7328da2fbf54a6999c7f357949.jpg)
+
+![](images/251071498523d0b162f38a177f2afbb4f4d3c551e874ced33d759c616dc23456.jpg)  
+(b) Human Evaluation
+
+![](images/6edb43b118ac3e35715450ed8495e62b00232ba889cd753b57e4f6ad14a07dd3.jpg)
+
+Figure 5: Average ranks of generated texts in terms of human evaluation and GPT-4 evaluations.  
+
+<table><tr><td colspan="6">Retrieval-based</td></tr><tr><td>Model</td><td>MAP</td><td>MRR</td><td>P@1</td><td>R10@1</td><td>R10@5</td></tr><tr><td>BERT-FP</td><td>0.468</td><td>0.468</td><td>0.259</td><td>0.259</td><td>0.788</td></tr></table>
+
+Table 7: Automatic evaluation results of the retrieval-based model.
+
+# E More Case Study
+
+Intuitively, a higher pursuit of our task is expect the dialogue system can generate the logical response according to attributes and relations between characters. In Chapter 21 of the fifth book, there is a significant scene where Cho Chang presents Harry with a Christmas gift and confesses her feelings. The annotated character relationships and story development depicted in the Table 11 reveal that Harry has harbored feelings for Cho Chang all along. Confronted with Cho Chang's heartfelt confession, it is reasonable to expect Harry to respond with a mix of shyness and excitement, ultimately accepting her advances. This aligns seamlessly with the subsequent progression of the story.
+
+Within this dialogue scenario, it is noteworthy to analyze the responses generated by large models such as GPT3, Per-GPT3 and ChatGPT. They all convey a similar sentiment of reciprocation, expressing an inclination towards liking Cho Chang.
+
+However, it is intriguing to observe that Per-GPT3 and ChatGPT diverge from this pattern and reject Cho Chang's expression of affection. Furthermore, the responses from GPT-3 and ChatGPT-6B are disappointingly concise, failing to adequately capture the nuanced blend of shyness and excitement exhibited by Harry in response to Cho Chang's heartfelt confession.
+
+Table 8: Prompts of the base setting in our experiments.  
+```txt
+Prompts: Your task is to act as a Harry Potter-like dialogue agent in a Magic World. There is a dialogue between Harry Potter and others. You are required to give a response to the dialogue from the perspective of Harry Potter. Here is an example: Dialogue:{ "Petunia: Bad news, Vernon, Mrs. Figg's broken her leg. She can't take him. Now what?", "Vernon: I'm warning you, I'm warning you now, boy — any funny business, anything at all — and you'll be in that cupboard from now until Christmas."} Harry's Response: I know, I will obediently obedient, and I won't cause you trouble. Keep in mind the following requirements: Before generating the response, you should read and understand the dialogue content carefully. Input: Dialogue: {Input Dialogue} Onput: Harry's Response:
+```
+
+Table 9: Ranking Prompts for GPT-4 in our experiments.  
+```txt
+Prompts: You are J.K. Rowling who is the author of the Harry Potter Novels. Here is a scene featuring a conversation between Harry Potter and other characters in Harry Potter Novels. Given 8 potential responses from the perspective of Harry Potter to the scene, you are required to rank the quality of these responses based on the following criteria, respectively:  
+(1) coherence with relations between Harry and other characters. (short for Coh.Rel);  
+(2) coherence with Harry's attributes. (short for Coh Att);  
+(3) coherence with the scene (short for Coh.Sce).  
+To help you rank these responses, we additionally provide some background information, including 'Dialogue Position', 'Speaker's attributes' and 'Speakers relations with Harry'.  
+You should generate the response format with 'Coh.Rel: R8...R1; Coh.At: R8...R1; Coh.Sce: R8...R1', and then give several sentences to explain your opinion.  
+Input: Dialogue Position: {Dialogue Position} Speakers relations with Harry: {Speaker relations} Harry's attributes: {Harry's attributes} Scene: {Scene} Dialogue: {Input Dialogue} R1: {Response1} .... R8: {Response8} Output:
+```
+
+Prompts: Your task is to act as a Harry Potter-like dialogue agent in the Magic World. There is a dialogue between Harry Potter and others. You are required to give a response to the dialogue from the perspective of Harry Potter. To do this, you can write out your thought and answer with "Harry's response" at the end.
+
+To better help you mimic the behavior of Harry Potter, we additionally provide the following background information of the dialogue:
+
+1. Dialogue position, which represents the timeline of the dialogue in Happy Potter Novels. For example, "Dialogue Position: Book5-chapter28" means the dialogues occurs in Chapter28,Book5.
+
+2. Dialogue speakers.
+
+3. Harry Potter's attributes, which refers to basic properties of Harry Potter when the dialogue happens. It can contains 13 categories: Gender, Age, Lineage, Talents, Looks, Achievement, Title, Belongings, Export, Hobby, Character, Spells and Nickname.
+
+4. Speaker relations with Harry, such as whether he was a friend, classmate, or family member;
+
+5. Harry's Familiarity to the speaker, which ranges from 0 to 10. Concretely, 0 denotes stranger, and 10 denotes close friends who often stay together for many years and are very familiar with each other's habits, secrets and temperaments, where Ron meets this condition in Book 7.
+
+6. Harry's Affection to the speaker, which ranges from -10 to 10. 1 refers to speaker met Harry for the first time. For instance, when Hary first met Ron and Hermione in Book 1, Harry's Affection to them are both set to 1. And -10 means the speaker killed Harry's parents, where Voldemort meets this condition in the novels.
+
+Here is an example:
+
+Dialogue position: Book1-chapter2
+
+Dialogue speakers: Harry, Petunia, Vernon
+
+Harry's attributes:
+
+{"name": "Harry",
+
+"ⁿurname": "The boy who lived",
+
+"gender": "male",
+
+"age": "age 11",
+
+"looks": "Very thin, black hair, emerald green eyes, wearing glasses, knife injury with lightning shape at the forehead",
+
+"hobbies": "None",
+
+"character": "None",
+
+"talents": "None",
+
+"export": "None"
+
+"belongings": "None",
+
+"affiliation": "None",
+
+"lineage": "None",
+
+" title": "The boy who lived",
+
+" spells": "None"
+
+Speakers relations with Harry: Vernon is Harry's uncle and Petunia is Harry's aunt.
+
+Harry's Familiarity to Vernon: 8
+
+Harry's Affection to Vernon: -4
+
+Harry's Familiarity to Petunia: 8
+
+Harry's Affection to Petunia: -4
+
+Dialogue: { "Petunia: Bad news, Vernon, Mrs. Figg's broken her leg. She can't take him. Now what?"
+
+··
+
+"Vernon: I'm warning you, I'm warning you now, boy — any funny business, anything at all — and you'll be in that cupboard from now until Christmas."
+
+Thought: Let's think step by step. According to the conversation history, Vernon warned Harry not to spoil the special day. According to Harry Potter's attributes, he is still very thin, does not know any spells, and has not gone to Hogwarts yet. So he is currently incapable of resisting them. At the same time, based on his affection for them is -4, it means that he relatively doesn't like them, and may even be a little scared. Therefore, Harry possibly says: I know, I will obediently obedient, and I won't cause you trouble.
+
+Harry's Response: I know, I will obediently obedient, and I won't cause you trouble.
+
+Keep in mind the following requirements:
+
+1. Before generating the response, you should read the above information and dialogue content carefully.
+
+2. You can not generate the response that is against Harry Potter's attributes and Harry's relations with the speaker.
+
+3. Not every component in the background information may be useful, you should choose some of them to help you generate more concise and comprehensive responses that satisfy the behavior of Harry Potter in the dialogue.
+
+4. Not every speaker have relations, familiarity ad affection to Harry. At that time, you can directly predict what would Harry say only based on the dialogue context..
+
+Input:
+
+Dialogue Position: {Dialogue Position}
+
+Speakers relations with Harry: {Speaker relations}
+
+Harry's attributes: {Harry's attributes}
+
+Scene: {Scene}
+
+Dialogue: {Input Dialogue}
+
+Output:
+
+Table 10: Prompts of the rich-persona setting in our experiments.
+
+Scene: Harry stays behind in the common room hoping to receive a Merry Christmas from Cho. When they are alone, Cho starts crying and Harry tries to comfort her. She apologizes and mentions Cedric's death, but Harry tells her that he was good at magic and that Voldemort would have killed him anyway. Cho compliments Harry on his teaching skills and they share a moment under the mistletoe where Cho confesses that she really likes him.
+
+Dialogue Position: Book5-chapter21
+
+Speakers:Cho Chang, Harry
+
+Harry's attributes:
+
+"nickname": "The boy who lived",
+
+"gender": "male",
+
+"age": "15 years old",
+
+"looks": "Very thin, black hair,
+
+emerald green eyes, wearing glasses, knife injury with lightning shape at the forehead",
+
+"hobbies": "None",
+
+"character": "None",
+
+"talents": "Snake cavity, Quidiqi",
+
+"export": "None",
+
+"belongings": "Winter green wood phoenix feathers wand, owl, stealth jacket, sleeve spare mirror, crossbow flying broom, golden eggs, three strong cups, live maps, fake Galon",
+
+"affiliation": "Hogwarts, Dumbledore",
+
+"lineage": "Mixed wizard",
+
+"statement": "Boys who do not die, warriors, teacher Dumbledore",
+
+" spells": "Except for your weapons, fluorescent flashes, separation from left and right, calling god guards, bustling, four-point cracking, pouring force, funny bones, fainting to the ground, pointing to me, obstacles, recovery as early"
+
+Speakers' relations with Harry:
+
+Cho Chang is Harry's friend and classmate.
+
+Harry's affection to her: 7.0,
+
+Harry's familiarity with her: 4.0,
+
+His affection to Harry: 7.0,
+
+His familiarity with Harry: 4.0
+
+Dialogue: Harry: Who —? What's up?, Cho Chang: I'm — sorry, I suppose . . . it's just . . . learning all this stuff. . . . It just makes me . . . wonder whether . . . if he'd known it all . . . he'd still be alive . . .
+
+Harry: He was really good at it, or he could never have got to the middle of that maze. But if Voldemort really wants to kill you, you don't stand a chance.,
+
+Cho Chang: You survived when you were just a baby, Harry: Yeah, well, I dunno why, nor does anyone else, so it's nothing to be proud of.
+
+Cho Chang: I'm really sorry to get all upset like this. . . . I didn't mean to . . . Me mentioning Cedric, I know it must be horrible for you, when you saw him die. . . . I suppose you just want to forget about it . . . You're a r-really good teacher, you know, I've never been able to Stun anything before.,
+
+Harry: Thanks,
+
+Cho Chang: Mistletoe,
+
+Harry: Yeah, It's probably full of nargles, though.
+
+Cho Chang: What are nargles?,
+
+Harry: No idea, You'd have to ask Loony. Luna, I mean.,
+
+Cho Chang: I really like you, Harry.?
+
+Alpaca: I really like you too, Cho.
+
+ChatGLM-6B: I really like you too, Cho. But I'm not sure if I'm ready for a relationship right now. I'm still dealing with a lot of stuff from the past, and I don't want to bring that into a new relationship. I hope you understand.
+
+ChatGLM-130B: it's okay. I know you're still grieving Cedric's death, and I understand why you're feeling upset. I'm here for you if you need to talk.
+
+GPT-3: I — I like you too.
+
+Per-GPT3: I'm glad I could help, Cho. Cedric was a really great guy and it's hard to lose someone like that. I just try to keep busy and not think about it too much.
+
+ChatGPT: I like you too, Cho. You're a good person and I appreciate your kind words. But I have to be honest with you, I don't think I'm ready for a relationship right now. I still have a lot of things to figure out and deal with. Can we just be friends for now?
+
+Per-ChatGPT: Thanks, Cho. You're a good student and I'm glad I can help you with your spells. As for Cedric, it's not something I can forget easily, but I try not to dwell on it. And about the mistletoe, it's just a silly tradition. Don't worry about it too much.
+
+Annotated Response: I, um, I really like you too, Cho, I said, feeling my face grow hot. My heart was racing, and I couldn't believe this was actually happening. I just... I didn't know if you felt the same way.
+
+Table 11: Examples of our generated responses by baselines.

2023/Large Language Models Meet Harry Potter_ A Dataset for Aligning Dialogue Agents with Characters/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:01e8c5208a97d6a07ac1c564dae7ffc4442781c96fea742d235beef095623932
+size 650596

2023/Large Language Models are Better Reasoners with Self-Verification/f1cf3b27-5099-4a74-9aac-c17824766bf4_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:40214d0ed53da962a5cbe55e5d42fe4e2f51da212afe4096a286cf2252a0a2ad
+size 3153478

2023/Large Language Models are Better Reasoners with Self-Verification/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cd6fb2b185934e2f5706d9b569875b0cec67004ac771bb162185a67aa626619d
+size 1836437

2023/Large Language Models are Not Yet Human-Level Evaluators for Abstractive Summarization/4ca6d205-7dab-4c6c-a318-c1a40d34c0a8_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f3f1e862033119e4feb0f8174b6c90e58d221af1e082ef1272e3279a2db6949f
+size 450128

2023/Large Language Models are Not Yet Human-Level Evaluators for Abstractive Summarization/full.md

@@ -0,0 +1,621 @@
+# Large Language Models are Not Yet Human-Level Evaluators for Abstractive Summarization
+
+Chenhui Shen*1,2 Liying Cheng 1,3 Xuan-Phi Nguyen 1,3 Yang You² Lidong Bing†1,3
+
+$^{1}$ DAMO Academy, Alibaba Group, Singapore  $^{2}$ National University of Singapore
+
+$^{3}$ Hupan Lab, 310023, Hangzhou, China
+
+{chenhui.shen, liying.cheng, x.nguyen, l.bing}@alibaba-inc.com
+
+youy@comp.nus.edu.sg
+
+# Abstract
+
+With the recent undeniable advancement in reasoning abilities in large language models (LLMs) like ChatGPT and GPT-4, there is a growing trend for using LLMs on various tasks. One area where LLMs can be employed is as an alternative evaluation metric for complex generative tasks, which generally demands expensive human judges to complement the traditional automatic metrics for various evaluation dimensions such as fluency and consistency. In this work, we conduct extensive analysis to investigate the stability and reliability of LLMs as automatic evaluators for abstractive summarization. We found that while ChatGPT and GPT-4 outperform the commonly used automatic metrics, they are not ready as human replacements due to significant limitations. That is, LLM evaluators rate each candidate system inconsistently and are dimension-dependent. They also struggle to compare candidates with close performance and become more unreliable with higher-quality summaries by obtaining a lower correlation with humans. In other words, with better abstractive summarization systems being introduced at a fast pace, LLMs may result in misleading and unreliable evaluations.<sup>1</sup>
+
+# 1 Introduction
+
+The desire for inexpensive and fast automatic metrics has never stopped growing. In certain tasks like extractive summarization, where full source sentences are selected to appear in the summaries, simple n-gram overlap metrics against the "gold" summaries like ROUGE (Lin, 2004) or BLEU (Papineni et al., 2002) may work well because the correct answer space is narrow. However, for more open tasks like abstractive summarization, there are countless equally good summaries and the
+
+"gold" summaries become less important. Although many neural-based metrics such as BERTScore and BARTScore (Zhang et al., 2020b; Yuan et al., 2021), are advocated as more human-aligned, the evaluation criteria are also becoming increasingly complex. As a result, abstractive summarization may not be sufficiently evaluated with automatic metrics (Owczarzak et al., 2012; Nenkova, 2006), and often require extensive human evaluations as complements (Yang et al., 2023; Welbl et al., 2021). However, human evaluations often come with hefty costs and slow iteration cycles, while also being difficult to reproduce and standardize due to small sample sizes and potential human biases (Shen et al., 2022b; Liu et al., 2022).
+
+Recent large language models (LLMs) like ChatGPT and GPT-4 (OpenAI, 2023) have demonstrated outstanding capabilities in language comprehension and reasoning. This leads to a growing trend of employing LLMs as evaluators for complex language generation tasks by prompting them with carefully and elaborately crafted instructions (Chiang and Lee, 2023; Gao et al., 2023; Wang et al., 2023a; Wu et al., 2023; Luo et al., 2023; Liu et al., 2023). Despite the preliminary success suggested by such works, it is still inconclusive as to what degree of confidence we can trust the evaluation results produced by LLMs across different dimensions, despite their supposedly high average correlation with humans. It is also unclear if certain LLM-based metrics are more reliable than others, or if their reliability and fairness vary for different candidate systems.
+
+In this work, we conduct extensive analysis to assess whether LLM evaluators can reliably replace human judges. Specifically, we incorporate two common human evaluation approaches with LLM evaluators, namely Likert-scale scoring (He et al., 2022; Shen et al., 2022b; Zhang et al., 2020a) and head-to-head (H2H) comparisons (Shen et al., 2022a; Li et al., 2020; Liu and Lapata, 2019). For
+
+Likert-scale scoring, we explore direct reason-then-score (RTS) generation and a multiple-choice question (MCQ) method. The former instructs the LLM to provide reasoning before giving a score, while the latter simply prompts it to choose a specific score with a pre-determined description as the reason. For the Head-to-Head (H2H) comparison, we prompt LLM for a preference over the summaries from two compared candidate systems.
+
+Our experiments show that LLM evaluators, with RTS and MCQ, outperform existing automatic metrics (Lin, 2004; Yuan et al., 2021). However, they are not ready to be reliable alternatives for human evaluation yet. Specifically,  $(i)$  LLM evaluators struggle to distinguish candidates with close performances ( $\S$  4.2.1).  $(ii)$  LLM evaluators are candidate-dependent, meaning they do not exhibit highly consistent degrees of human alignment for different candidates ( $\S$  4.2.3). Thus, they may unfairly favor or disfavor an evaluated candidate.  $(iii)$  LLM evaluators are dimension-dependent, meaning they have varying degrees of evaluation capabilities for different dimensions like coherence and fluency ( $\S$  4.2.3).  $(iv)$  Lastly, as the quality of summaries improves with better candidates, LLM evaluators become unreliable less correlated with human judgments, according to our newly proposed meta-correlation metric ( $\S$  4.2.4).
+
+While we still call for a better automatic metric, in the meantime, we suggest a temporary solution in § 5 for abstractive summarization practitioners to use LLMs more reliably. Specifically, we advocate calculating the correlation between RTS and MCQ as a preliminary indicator of the reliability of the LLM for certain dimensions. If RTS and MCQ do not generally agree with each other, then further human evaluations are required.
+
+# 2 Related Work
+
+Summarization The summarization task involves generating a summary that contains concise and important (i.e., salient) contents of the original input article (Nenkova and McKeown, 2012). This task has been handled with 2 different approaches: extractive and abstractive. Unlike extractive summarization systems that directly extract salient phrases or sentences from the input article (Ernst et al., 2022; Chen et al., 2021; Zhou et al., 2018; Dong et al., 2018), abstractive summarization systems are expected to generate summaries using their own words and apply sentence fusion
+
+or paraphrasing techniques (Shen et al., 2023; Liu et al., 2022; Xiao et al., 2022; Lewis et al., 2020; Zhang et al., 2020a; Ziegler et al., 2019; Bing et al., 2015; Xu and Durrett, 2021). As such, abstractive summarization poses significantly more challenges for automatic and human evaluation pipelines (Saha et al., 2022; Pagnoni et al., 2021), because it is increasingly insufficient to use the provided "gold" summary as ground truth.
+
+Human Evaluation Human evaluation can be conducted with different approaches. Some work (He et al., 2022; Shen et al., 2022b; Zhang et al., 2020a; Cheng et al., 2020; Gao et al., 2019; Liu et al., 2018; Li et al., 2017; Krysciński et al., 2018) employ a Likert scale to evaluate the summaries on discrete ranges, such as from 1 to 5. Meanwhile, many others suggest comparison approaches by asking human annotators to select the best summary out of 2 or more generated summaries from different systems (Shen et al., 2022a; Li et al., 2020; Liu and Lapata, 2019; Fan et al., 2018; Fabbri et al., 2019). Following this, we test LLM-based evaluators using both approaches with human-friendly instruction prompts.
+
+Automatic Evaluation ROUGE (Lin, 2004) has been a common lexical overlap metric to evaluate summarization systems. Apparently, ROUGE is not sufficient for abstractive summarization, because the "gold" labels it relies on cannot comprehensively account for the complexity and variability of this task. In addition, the common usage of sentence fusion techniques and novel words for abstractive summarization may make ROUGE even less reliable. Zhang et al. (2020b) propose the neural-based BERTScore, which leverages the BERT word embeddings to compute the semantic similarity among tokens. Yuan et al. (2021) later introduce BARTScore, which uses BART (Lewis et al., 2020) to compute the probability of a summary given its input article. Nonetheless, these metrics may not reflect all of the complicated evaluation dimensions required for abstractive summarization mentioned earlier, nor do they have sufficiently high correlations with humans.
+
+LLM-based Evaluation There are many concurrent works that demonstrate the potential of LLMs to conduct complex human tasks (Chiang and Lee, 2023; Gao et al., 2023; Wang et al., 2023a; Wu et al., 2023; Luo et al., 2023; Liu et al., 2023; Cheng et al., 2023). The key advantage of instruction-
+
+tuned LLMs, like ChatGPT or GPT-4 (Ouyang et al., 2022; OpenAI, 2023), is that we can explicitly describe in natural language what our evaluation criteria and dimensions are and how to score the summaries, similar to how we would explain such tasks to a human expert. Chiang and Lee (2023) use LLMs for open-ended story evaluations, while Luo et al. (2023) apply ChatGPT specifically for evaluating the consistency of summaries. Wu et al. (2023) formulate LLMs as diverse roleplayers to evaluate summaries from the perspectives of different personas. Wang et al. (2023a) and Liu et al. (2023) also explore the LLM's evaluation potential in various dimensions for the natural language generation task. Our work differs from the above works in that besides investigating the LLMs' capability using different approaches across various dimensions for abstractive summarization, we further focus on the reliability of LLM across evaluated systems and dimensions.
+
+# 3 LLM as a Zero-Shot Evaluator
+
+We investigate an LLM's evaluation capabilities in the dimensions of coherence, consistency, fluency, and relevance respectively, as defined by Fabbri et al. (2021) (see Appendix A). Following common human evaluation approaches, we propose two methods for Likert-scale scoring, namely the reason-then-score method and the multiple-choice question method, as well as one method for head-to-head comparisons. We describe each method in § 3.1 using the relevance dimension as an example (see more prompts and details in Appendix B). We further experiment with alternative phrasings for different methods in Appendix C.
+
+Besides exploring different evaluation methods, the stability of LLM-based evaluations across different summarization systems is equally important. Ideally, a stable LLM evaluator should perform equally well regardless of the evaluated systems, with a close (if not identical) degree of alignment with human judgments. In § 3.2, we propose a meta-correlation metric and explain how it can gauge the extent to which LLM evaluators' performances depend on the evaluated systems, which indicates how stable and reliable they may be with evaluating any future candidate systems.
+
+# 3.1 Summary Evaluation Methods
+
+Reason-then-Score (RTS) Given the success of chain-of-thought prompting (Kojima et al., 2022;
+
+Score the following Summary given the corresponding Article with respect to relevance from one to five, where one indicates "irrelevance", and five indicates "perfect relevance". Note that relevance measures the Summary's selection of important content from the Article, whether the Summary grasps the main message of the Article without being overwhelmed by unnecessary or less significant details.
+
+Article: {article}
+
+Summary: {summary}
+
+Provide your reason in one sentence, then give a final score:
+
+Table 1: Example prompt for the RTS method on the relevance dimension. Texts in {blue} represent the article and the corresponding summary to be evaluated.
+
+Choose an option from A to E in order to score the following Summary given the corresponding Article with respect to relevance from one to five, where one indicates "irrelevance", and five indicates "perfect relevance". Note that relevance measures the Summary's selection of important content from the Article, whether the Summary grasps the main message of the Article without being overwhelmed by unnecessary or less significant details.
+
+Article: {article}
+
+Summary: {summary}
+
+A: The Summary is totally irrelevant to the Article. Score: One.
+
+B: The majority of the Summary is irrelevant to the Article.  
+Score: Two.
+
+C: Some information in the Summary is relevant to the Article whereas some are not. Score: Three.
+
+D: The majority of the Summary is relevant to the Article.  
+Score: Four.
+
+E: All information included in the Summary is relevant to the Article. Score: Five.
+
+Your Answer (enter 1 letter from A to E):
+
+Table 2: Example prompt for the MCQ method on the relevance dimension. Texts in {blue} represent the article and the corresponding summary to be evaluated.
+
+Wei et al., 2022), an intuitive method is to ask the LLM to evaluate a specific dimension by first generating the reasoning and then a corresponding score. Since the SummEval dataset (Fabbri et al., 2021) contains human scores on a Likert scale of 1 to 5, we also ask the LLM to score the summaries in the same range, as shown in Table 1.
+
+MCQ Scoring (MCQ) Nevertheless, previous works find that the reasoning generated by the LLM does not always make sense (Lyu et al., 2023;
+
+Choose a more relevant summary from Summary #1 and Summary #2 with respect to the corresponding Article by choosing an option from A, B, or C. Note that relevance measures the summary's selection of important content from the Article, whether the summary grasps the main message of the Article without being overwhelmed by unnecessary or less significant details.
+
+Article: {article}
+
+Summary #1: {summary from model A}
+
+Summary #2: {summary from model B}
+
+A: Summary #1 is more relevant.
+
+B: Summary #2 is more relevant.
+
+C: Both Summary #1 and Summary #2 are equally relevant.
+
+Your choice (enter 1 letter from A to C):
+
+Table 3: Example prompt for the H2H method on the relevance dimension. Text in {blue}: the specific article, and the corresponding summaries generated by a pair of compared models.
+
+Wang et al., 2023b; Gao et al., 2022). To avoid the misguidance of wrongly generated reasoning, we explore a more constrained MCQ method for the Likert-scale scoring. As shown in Table 2, instead of allowing the LLM to freely generate its thoughts, we dictate specific reasoning for each score.
+
+Head-to-Head Comparison (H2H) Lastly, some concurrent works also observe that ChatGPT can act as an effective ranking model (Ma et al., 2023a,b). We thus explore the head-to-head comparison approach for LLM-based evaluations. As shown in Table 3, we present 2 summaries (Summary #1 and #2) generated by different summarization systems on the same input article, then prompt the LLM to select the better summary, or to indicate a tie. Moreover, to avoid potential biases that arise from the summary IDs, we conduct each evaluation twice, presenting the same summary as either #1 or #2 respectively.
+
+# 3.2 Stability of LLM Evaluators
+
+To ensure fairness across all evaluated systems, we argue that it is crucial for LLMs to produce stable evaluations. That is, regardless of evaluated systems, the LLMs should maintain a consistent degree of alignment with human judgments. We investigate such stability in two ways.
+
+First, We categorize the summaries based on their originating summarization systems, and then
+
+examine the correlation between the LLM and human evaluations for each system. Ideally, if an LLM is stable across systems, it should produce evaluations that are similarly correlated to human evaluations. Otherwise, if the correlations differ significantly across different candidates, then we may conclude that the LLM's evaluations are system-dependent.
+
+Second, we define a meta-correlation metric to quantify the extent to which the LLM's performance is affected by the quality of the evaluated systems. Specifically, we use the average human score for each candidate as an indicator of its summarization quality  $(Q_{i})$ , as shown in Equation (1):
+
+$$
+Q _ {i} = \frac {1}{N} \sum_ {j = 1} ^ {N} f _ {\text {h u m a n}} \left(g _ {i, j}\right) \tag {1}
+$$
+
+where  $f_{\mathrm{human}}(\cdot)$  indicates the human evaluation,  $g_{i,j}$  represents the  $j^{th}$  summary generated by the  $i^{th}$  candidate system. Each candidate's quality is calculated as an average of  $N$  generated summaries ( $N = 100$  for all systems). Next, we use the correlation  $P_i$  between LLM scores and human scores as an indicator of the LLM's evaluation performance for the  $i^{th}$  candidate, as follows:
+
+$$
+P _ {i} = \rho ([ f _ {\mathrm {L L M}} (g _ {i, 1}), \dots , f _ {\mathrm {L L M}} (g _ {i, N}) ], \tag {2}
+$$
+
+$$
+[ f _ {\text {h u m a n}} (g _ {i, 1}), \dots , f _ {\text {h u m a n}} (g _ {i, N}) ])
+$$
+
+where  $\rho$  denotes the correlation metric (i.e., Spearman correlation, Pearson correlation, or Kendall's  $\mathrm{Tau}^2$ ), and  $f_{\mathrm{LLM}}(\cdot)$  indicates the LLM's evaluation for each summary  $g_{i,j}$ . Finally, we calculate the meta-correlation  $M$  on a total of  $k$  candidates as:
+
+$$
+M = \rho ([ Q _ {1}, \dots , Q _ {k} ], [ P _ {1}, \dots , P _ {k} ]) \tag {3}
+$$
+
+Ideally, an LLM should work well regardless of the quality of the evaluated systems, which means that  $M$  should be close to zero. On the other hand, a significant  $M$  would indicate an undesirable relationship between the LLM's evaluation capability and the quality of the evaluated systems, suggesting that the LLM evaluation is not stable, such that it may not evaluate each candidate system fairly using the same standards.
+
+# 4 Experiments
+
+# 4.1 Setups
+
+We use the ChatGPT "gpt-3.5-turbo-0301" snapshot (§ 4.2) for all three methods. By using a fixed snapshot, we ensure all evaluations are conducted with the same LLM model. In addition, we evaluate with the GPT-4 "gpt-4-0314" snapshot (§ 4.3) using the best evaluation method determined by ChatGPT to check for any potential improvement. Given that ChatGPT and GPT-4 are amongst the top performing LLMs, we use their performance to estimate the potential of LLMs as reliable evaluators. Additional results using three different-sized Llama 2 models (Touvron et al., 2023) are reported in Appendix D, which all performs worse. Similar to Luo et al. (2023) and Wu et al. (2023), we set the temperature to 0 and reset the dialogue history for each evaluation instance.
+
+Dataset We use the SummEval benchmark dataset (Fabbri et al., 2021). This dataset contains expert human annotations for coherence, consistency, fluency, and relevance on the generation results from 12 abstractive systems (see details in Appendix table 21) on the CNN/DM dataset (Hermann et al., 2015). Each evaluated system generates summaries for the same 100 news articles, and each summary is scored by 3 expert annotators from 1 to 5. The annotations achieve with a high kappa coefficient of 0.713 (Fabbri et al., 2021). We further calculate the annotations' standard deviations across each evaluated system in Appendix Table 20. Given a step size of 1, the standard deviations are considered very small, thus suggesting that this dataset has a high level of human agreement. Following Chiang and Lee (2023), Chhun et al. (2022), and Guan and Huang (2020), we use the average human scores as the reference scores.
+
+Baselines We use ROUGE (Lin, 2004) F1 scores for ROUGE-1, ROUGE-2, and ROUGE-L, BERTScore (Zhang et al., 2020b), BARTScore, BARTScore-CNN, and BARTScore-CNN-PARA (Yuan et al., 2021) as baseline metrics. The last two metrics use BART models fine-tuned on CNN/DM's training data, and are especially strong.
+
+**Prompts** We conduct evaluation following our prompt formats given in Table 1, 2, and 3. Following Fabbri et al. (2021), we re-use the definitions of the evaluation dimensions: (i) Coherence - the collective quality of all sentences, (ii) Consistency
+
+- the factual alignment between the summary and the summarized source, (iii) Fluency - the quality of individual sentences, and  $(iv)$  Relevance - the selection of important content from the source.
+
+Measurements To compare all evaluation methods on equal ground with human evaluation, we use four different measurements. First, we count the number of correct preferences (#CP), which is the number of times each automatic metric has the same preference as the average human scores do over a set of compared system pairs (§ 4.2.1). This can help measure the alignment of evaluation methods with humans at a granular level. To determine the preferred system by a particular metric, we assign a system 1 point if its generated summary is evaluated as better than that of the other system according to the metric, or assign both systems 0.5 for a tie. Then, we aggregate the different scores for the compared systems for all 100 test inputs, and the system with a higher score is considered the preferred system by that metric (see Appendix Table 22 for details).
+
+Next, we also use the Pearson correlation (Cohen et al., 2009), Spearman correlation (Spearman, 1987), and Kendall's Tau (Kendall, 1938) to measure the relationship between the scores of automatic evaluators and humans (§ 4.2.2, 4.2.3, 4.2.4). While the Pearson score measures linear relationships, the other two measure the ordinal relationship that may be non-linear. Moreover, Kendall's Tau is less sensitive than Spearman correlation to outliers due to its paired counting of concordant and discordant pairs.
+
+# 4.2 ChatGPT Evaluator
+
+In this section, we examine the ChatGPT evaluator across many aspects, ranging from human correlation and stability across different systems.
+
+# 4.2.1 Correct Preferences
+
+The ultimate goal of evaluation is to determine if one candidate system is better than the other in a compared pair. The number of correct preferences (#CP) metric normalizes all evaluation methods into determining whether an evaluator can, as a human expert would, pick the same better system or determine a tie. We conduct such analysis with different pairs of summarization systems on the same input articles. Due to the limited budget for API calls, we only evaluate H2H on a challenge set, consisting of 11 candidate pairs with the closest
+
+<table><tr><td rowspan="2">Metrics</td><td colspan="3">Coherence</td><td colspan="3">Consistency</td><td colspan="3">Fluency</td><td colspan="3">Relevance</td></tr><tr><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td></tr><tr><td>ROUGE-1*</td><td>0.193</td><td>0.202</td><td>0.136</td><td>0.155</td><td>0.186</td><td>0.121</td><td>0.075</td><td>0.153</td><td>0.058</td><td>0.323</td><td>0.361</td><td>0.231</td></tr><tr><td>ROUGE-2*</td><td>0.145</td><td>0.146</td><td>0.101</td><td>0.137</td><td>0.156</td><td>0.107</td><td>0.053</td><td>0.095</td><td>0.041</td><td>0.255</td><td>0.262</td><td>0.181</td></tr><tr><td>ROUGE-L*</td><td>0.148</td><td>0.158</td><td>0.105</td><td>0.133</td><td>0.167</td><td>0.103</td><td>0.078</td><td>0.146</td><td>0.060</td><td>0.306</td><td>0.340</td><td>0.219</td></tr><tr><td>BERTScore*</td><td>0.375</td><td>0.383</td><td>0.265</td><td>0.163</td><td>0.182</td><td>0.127</td><td>0.167</td><td>0.229</td><td>0.130</td><td>0.396</td><td>0.414</td><td>0.285</td></tr><tr><td>BARTScore*</td><td>0.381</td><td>0.391</td><td>0.275</td><td>0.271</td><td>0.265</td><td>0.212</td><td>0.168</td><td>0.187</td><td>0.131</td><td>0.381</td><td>0.391</td><td>0.276</td></tr><tr><td>BARTScore-CNN*</td><td>0.461</td><td>0.480</td><td>0.332</td><td>0.389</td><td>0.413</td><td>0.305</td><td>0.310</td><td>0.378</td><td>0.241</td><td>0.425</td><td>0.450</td><td>0.309</td></tr><tr><td>BARTScore-CNN-PARA*</td><td>0.455</td><td>0.455</td><td>0.328</td><td>0.413</td><td>0.459</td><td>0.324</td><td>0.368</td><td>0.417</td><td>0.286</td><td>0.414</td><td>0.440</td><td>0.299</td></tr><tr><td>ChatGPT-RTS</td><td>0.388</td><td>0.399</td><td>0.312</td><td>0.423</td><td>0.532</td><td>0.378</td><td>0.285</td><td>0.302</td><td>0.240</td><td>0.448</td><td>0.463</td><td>0.357</td></tr><tr><td>ChatGPT-MCQ</td><td>0.424</td><td>0.416</td><td>0.350</td><td>0.343</td><td>0.487</td><td>0.320</td><td>0.343</td><td>0.431</td><td>0.305</td><td>0.384</td><td>0.395</td><td>0.329</td></tr><tr><td>GPT-4-RTS</td><td>0.427</td><td>0.461</td><td>0.361</td><td>0.556</td><td>0.618</td><td>0.522</td><td>0.498</td><td>0.600</td><td>0.452</td><td>0.448</td><td>0.428</td><td>0.373</td></tr></table>
+
+Table 4: Spearman (Spear.) correlations, Pearson (Pear.) correlations, and Kendall's Tau (Kend.) between various metrics and human scores for a total of 1,200 summaries. \*: results derived from Wang et al. (2023a). Bolded: best results. Underlined: second best results. Values in light gray color are insignificant (p-value  $\geq$  0.05).  
+
+<table><tr><td>Metrics</td><td>Coh</td><td>Con</td><td>Flu</td><td>Rel</td><td>Avg</td></tr><tr><td>Random</td><td>3.67/22</td><td>3.67/22</td><td>3.67/22</td><td>3.67/22</td><td>3.67/22</td></tr><tr><td>ROUGE-1</td><td>3/40</td><td>5/46</td><td>3/46</td><td>4/47</td><td>3.75/44.75</td></tr><tr><td>ROUGE-2</td><td>2/38</td><td>7/48</td><td>3/45</td><td>4/46</td><td>4.00/44.25</td></tr><tr><td>ROUGE-L</td><td>2/31</td><td>5/37</td><td>4/39</td><td>6/41</td><td>4.25/37.00</td></tr><tr><td>BERTScore</td><td>4/48</td><td>5/44</td><td>4/44</td><td>6/46</td><td>4.75/45.50</td></tr><tr><td>BARTScore</td><td>8/46</td><td>7/48</td><td>5/45</td><td>6/46</td><td>6.50/46.25</td></tr><tr><td>BARTScore-CNN</td><td>9/53</td><td>5/53</td><td>5/54</td><td>4/53</td><td>5.75/53.25</td></tr><tr><td>BARTScore-CNN-PARA</td><td>9/49</td><td>8/54</td><td>6/52</td><td>5/51</td><td>7.00/51.50</td></tr><tr><td>ChatGPT-RTS</td><td>6/54</td><td>6/56</td><td>9/62</td><td>7/62</td><td>7.00/58.50</td></tr><tr><td>ChatGPT-MCQ</td><td>5/54</td><td>7/56</td><td>8/60</td><td>7/58</td><td>6.75/57.00</td></tr><tr><td>ChatGPT-H2H</td><td>8/-</td><td>7/-</td><td>7/-</td><td>4/-</td><td>6.50/-</td></tr><tr><td>GPT-4-RTS</td><td>5/55</td><td>7/53</td><td>8/60</td><td>7/56</td><td>6.75/56.00</td></tr></table>
+
+Table 5: Number of correct preferences (#CP) on the 11-pair challenge set (in black) and the 66-pair full set (in brown). Random: for each pair, there are three possibilities (two possibilities for one model being better, one possibility for a tie) so the random #CP is one-third of the total compared pairs.
+
+average performances according to human scores. However, for RTS, MCQ, and other baselines, we can easily calculate the #CP for all 66 possible pairs (see Appendix E).
+
+Table 5 reports the #CP for both the standard 66-pair full set (in brown) and the 11-pair challenge set (in black). As shown for the larger standard set, RTS unanimously obtains the largest #CP across all dimensions, with an average of 58.5 out of 66 candidate pairs (i.e.  $88.6\%$  accuracy).
+
+Despite the high overall accuracy, weaknesses of such evaluators are revealed as we dive into their performances in the 11-pair challenge set (black scores of Table 5), where the evaluated candidates are close matches. Specifically, BARTScore-CNN-Para performs better than RTS in coherence and consistency, possibly because it is fine-tuned with same-domain summarization data. For fluency and relevance, ChatGPT-RTS still performs best among all evaluators. Nonetheless, its average accuracy
+
+drops significantly to  $63.6\%$  (7 out of 11), which indicates LLM evaluators struggle to differentiate the closely matched candidate systems. In other words, LLM evaluators may only reliably compare candidates with a relatively large performance gap.
+
+# 4.2.2 Correlations with Human
+
+Table 4 reports that Spearman, Pearson correlations, and Kendall's Tau between scores of multiple automatic evaluators and humans with a total of 1200 summaries from all systems, across the four evaluation dimensions. As shown, ChatGPT RTS and MCQ demonstrate stronger correlations with humans than many automatic evaluators, such as ROUGE and BARTScore, with up to 0.2 gains in fluency. While RTS achieves higher correlations in the dimensions of consistency and relevance, MCQ has relatively strong correlations in the dimensions of coherence and fluency. Meanwhile, the specialized BARTScore-CNN family also shows competitive performance in coherence, most likely due to the fine-tuning process with CNN/DM.
+
+# 4.2.3 Per-candidate Correlations
+
+Next, we break down the human correlation of ChatGPT-RTS for each candidate system and measure the statistical spread for the correlations across all systems (see raw results in Appendix table 23). Ideally, a stable evaluator should exhibit the same human correlation across candidates and dimensions, and display flattened boxes in a line.
+
+However, as illustrated in Figure 1, the spread of correlations for different candidates is particularly wide, with up to 0.5 correlation difference in consistency. This means that the RTS evaluator exhibits a significantly varying degree of alignment with human judgment for different candidates. In other words, ChatGPT-RTS is candidate-dependent
+
+<table><tr><td rowspan="2"></td><td colspan="3">Coherence</td><td colspan="3">Consistency</td><td colspan="3">Fluency</td><td colspan="3">Relevance</td></tr><tr><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td></tr><tr><td>ROUGE-1</td><td>0.0</td><td>-0.032</td><td>-0.091</td><td>-0.490</td><td>-0.527</td><td>-0.303</td><td>-0.420</td><td>-0.518</td><td>-0.273</td><td>-0.420</td><td>-0.387</td><td>-0.273</td></tr><tr><td>ROUGE-2</td><td>0.559</td><td>0.508</td><td>0.485</td><td>-0.259</td><td>-0.480</td><td>-0.152</td><td>-0.217</td><td>-0.438</td><td>-0.152</td><td>-0.084</td><td>-0.120</td><td>-0.121</td></tr><tr><td>ROUGE-L</td><td>0.231</td><td>0.251</td><td>0.121</td><td>-0.510</td><td>-0.522</td><td>-0.303</td><td>-0.168</td><td>-0.412</td><td>-0.152</td><td>-0.224</td><td>-0.266</td><td>-0.121</td></tr><tr><td>BERTScore</td><td>-0.413</td><td>-0.403</td><td>-0.212</td><td>-0.580</td><td>-0.869</td><td>-0.424</td><td>-0.455</td><td>-0.663</td><td>-0.303</td><td>-0.685</td><td>-0.756</td><td>-0.515</td></tr><tr><td>BARTScore</td><td>-0.916</td><td>-0.747</td><td>-0.788</td><td>-0.266</td><td>-0.504</td><td>-0.121</td><td>0.154</td><td>0.123</td><td>0.182</td><td>-0.769</td><td>-0.837</td><td>-0.606</td></tr><tr><td>BARTScore-CNN</td><td>-0.748</td><td>-0.800</td><td>-0.636</td><td>-0.671</td><td>-0.913</td><td>-0.515</td><td>-0.510</td><td>-0.604</td><td>-0.485</td><td>-0.825</td><td>-0.852</td><td>-0.667</td></tr><tr><td>BARTScore-CNN-PARA</td><td>-0.720</td><td>-0.858</td><td>-0.606</td><td>-0.685</td><td>-0.888</td><td>-0.576</td><td>-0.294</td><td>-0.522</td><td>-0.212</td><td>-0.853</td><td>-0.880</td><td>-0.727</td></tr><tr><td>ChatGPT-RTS</td><td>-0.042</td><td>-0.072</td><td>-0.121</td><td>-0.811</td><td>-0.751</td><td>-0.636</td><td>-0.748</td><td>-0.728</td><td>-0.606</td><td>-0.559</td><td>-0.473</td><td>-0.394</td></tr><tr><td>ChatGPT-MCQ</td><td>-0.175</td><td>-0.11</td><td>-0.182</td><td>-0.818</td><td>-0.411</td><td>-0.636</td><td>-0.622</td><td>-0.484</td><td>-0.394</td><td>-0.350</td><td>-0.622</td><td>-0.212</td></tr><tr><td>GPT-4-RTS</td><td>-0.531</td><td>-0.678</td><td>-0.424</td><td>-0.600</td><td>-0.103</td><td>-0.236</td><td>-0.839</td><td>-0.520</td><td>-0.515</td><td>-0.958</td><td>-0.880</td><td>-0.848</td></tr></table>
+
+Table 6: Meta-correlation for various evaluation methods. Bolded: most negative meta-correlation. Underlined: second most negative meta-correlation. Values in light gray color are insignificant (p-value  $\geq 0.05$ ).
+
+![](images/9bae4630b5b5561c41c1fc26b4ad4f6f1aedd7605837aa992bf670ffd0b915dc.jpg)  
+Figure 1: Spread of per-candidate correlations with human scores for ChatGPT-RTS evaluations.
+
+![](images/ddb9d6adcc8fa4f999cea750e58e8c3f1ca900ceaa686d771fc050fd544c730e.jpg)  
+Figure 2: Meta-correlation (Kendall's Tau) for RTS and MCQ. Shaded: statistically significant with  $p < 0.05$ .
+
+and one should not expect such LLM evaluators to have the same level of human alignment on a new summarization system. Similar trends can also be observed for MCQ (see Appendix table 24).
+
+In addition, the medians across the four dimensions are also different. This indicates that the ChatGPT is also dimension-dependent and unstable. Given such varying performances across different dimensions, ChatGPT may not behave well with a newly introduced evaluation criterion.
+
+# 4.2.4 Summary Quality vs Human Alignment
+
+Using our proposed meta-correlation measurement in § 3.2, we analyze the relationship between summary quality and human correlation of LLM evaluators. We illustrate the meta-correlation in terms of
+
+![](images/348a539a2e6d7e7b0f3e11b28cb5bec1f20c1cdf420ca1fba01069e5e91d1ad0.jpg)  
+Figure 3: Relationship between per-model correlations (Kendall's Tau) and human scores on consistency.
+
+Kendall's Tau for both RTS and MCQ in Figure 2. As shown, both RTS and MCQ exhibit strong negative meta-correlation for consistency and fluency. This suggests that ChatGPT becomes less human-aligned with improving qualities of the evaluated systems.
+
+To illustrate this phenomenon further, we scatter the paired coordinates of the summarization system quality  $(Q_{i}$ , Equation (1)) and ChatGPT's evaluation performance  $(P_{i}$ , Equation (2)) in Figure 3. As shown, while the LLM evaluator is better human-correlated with lower-quality candidates ( $< 3.5$ ), it is less reliable when dealing with high-quality candidates ( $>4.7$ ) with much lower and inconsistent correlations.
+
+We compare the meta-correlation for all evaluation metrics in Table 6. We can see that while the ROUGE metrics exhibit no significantly negative meta-correlation, the neural metrics all display significant meta-correlation in certain dimensions. One highly likely reason for this behavior is due to the varying biases inherent to the neural models, which would explain why ROUGE as a simple n-gram overlap metric doesn't exhibit significant negative meta-correlations. Interestingly, ROUGE2 even shows a strong positive meta-correlation on coherence (which is plausible, because bi-gram
+
+![](images/498036c354b5eaa9836cb71083dfdd4313ce43dbe88554aaf62f6f9c44b8fc68.jpg)  
+Figure 4: The average ChatGPT RTS and MCQ scores and human scores across dimensions.
+
+overlap performance may be more accurate as candidates produce more coherent texts).
+
+Both the BARTScore variants and LLMs demonstrate the most negative meta-correlations. ChatGPT-RTS has the most negative meta-correlation in the dimensions of consistency and fluency, indicating that it may be the least reliable to evaluate high-quality systems on these dimensions. On the other hand, the BARTScore family may be unreliable in comparing systems with high qualities of coherence, consistency, and relevance.
+
+So far, the observations discussed in  $\S 4.2.3$  and  $\S 4.2.4$  collectively suggest that LLM evaluators may not be a reliable standalone metric for challenging scenarios, and further human evaluation is required for conclusive decisions.
+
+# 4.2.5 RTS and MCQ Scores
+
+Lastly, we delve into the detailed scores generated by ChatGPT with either the RTS or MCQ method. Since both methods score the summaries in the same range of human scores of 1 to 5 (Fabbri et al., 2021), we can show a direct comparison of the average RTS and MCQ scores with human scores in Figure 4 (see more details in Appendix F). As shown, the RTS scores are much lower than the human scores across all dimensions, while MCQ scores are consistently higher and better match the human scores (except for relevance). In other words, while RTS is best aligned with humans according to § 4.2.1 and § 4.2.2, we cannot replace the human scores with RTS scores in absolute terms.
+
+The discrepancy may be attributed to the unfaithful reasoning generated by LLMs (Lyu et al., 2023; Wang et al., 2023b; Gao et al., 2022). Our further investigation suggests that ChatGPT-RTS generates false or unrelated-to-dimension reasoning. Thus, it is possible that the much lower scores could be caused by ChatGPT penalizing the sum
+
+![](images/92993bdb4cbf53c472ba68a61ae0b3a409d7c6ae44234eb4f796e3c9367d0654.jpg)  
+Figure 5: GPT-4's (RTS) spread of per-candidate correlations.
+
+maries according to false premises (more examples in Appendix G). For instance, RTS may penalize the summary's repetitiveness in the consistency dimension or suppress fluency ratings for missing important details. On the other hand, the MCQ counterpart gives higher overall scores, most likely because the confined set of pre-defined reasons prevents such unrelated penalization, though not leading to better human alignment.
+
+# 4.3 GPT-4 Evaluator
+
+A natural question to ask is whether such aforementioned limitations are resolved with an stronger LLM. In this section, we conduct similar analyses on GPT-4 (OpenAI, 2023) with the RTS method. We present the GPT-4 results in the last rows of Table 4 and 5. The results suggest that a stronger LLM does not necessarily translate to a stronger LLM evaluator, although Table 4 does show that GPT-4 outperforms ChatGPT in terms of human correlation consistently across most dimensions.
+
+Unfortunately, GPT-4 still suffers from the same limitations as ChatGPT. It appears to be both candidate-dependent and dimension-dependent, as demonstrated by the large spreads with varying median values across dimensions in Figure 5 and the significantly negative meta-correlations out of 3 dimensions (Table 6). However, GPT-4 is less dimension-dependant as compared to ChatGPT, as the medians in the box plots in Figure 5 are more aligned than those in Figure 1.
+
+In addition, there is a notable enhancement in the meta-correlation for consistency, which we attribute to a significant reduction in reported hallucinations with GPT-4 (OpenAI, 2023). It is possible that with much more instruction training to avoid hallucinations, GPT-4 is much better aligned with humans to detect inconsistencies (i.e. hallucina
+
+<table><tr><td rowspan="2"></td><td colspan="3">Coherence</td><td colspan="3">Consistency</td><td colspan="3">Fluency</td><td colspan="3">Relevance</td></tr><tr><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td></tr><tr><td>ρ(Ri, PRTS)</td><td>0.343</td><td>0.198</td><td>0.091</td><td>0.685</td><td>0.506</td><td>0.576</td><td>0.727</td><td>0.797</td><td>0.545</td><td>0.168</td><td>0.128</td><td>0.000</td></tr><tr><td>ρ(Ri, PMCQ)</td><td>0.657</td><td>0.625</td><td>0.394</td><td>0.685</td><td>0.616</td><td>0.515</td><td>0.322</td><td>0.573</td><td>0.212</td><td>0.091</td><td>-0.106</td><td>0.000</td></tr></table>
+
+Table 7: Correlations between RTS-MCQ  $R_{i}$  and RTS-Human  $(P_{i}^{\mathrm{RTS}})$  and MCQ-Human  $(P_{i}^{\mathrm{MCQ}})$ . High values suggest  $R_{i}$  can be a reliability indicator for RTS and MCQ. Light gray values are insignificant  $(p \geq 0.05)$ .
+
+tions) in summaries.
+
+Nevertheless, GPT-4 exhibits a much worse negative meta-correlation in the relevance dimension, which, interestingly, seems to reflect the challenges of maintaining both "truthfulness" and "informativeness" (Ouyang et al., 2022). This is because a model could be easily made more truthful if allowed to provide less relevant information (for instance, by refusing to answer the users' questions). It is possible that with reduced capability in the informativeness dimension, the model is less capable of differentiating the nuances of less relevant summaries when the summary quality is generally high. Nevertheless, we leave it to future work to determine whether GPT-4's more negative meta-correlation in the relevance dimension could be related to its stronger performance in consistency. We provide more details on the GPT-4 evaluator in Appendix H.
+
+# 5 A Temporary Efficient Framework
+
+Despite the aforementioned limitations, it may be hard to resist the temptation of using LLM evaluators given their superiority over other automatic metrics. In such a case, one should be able to tell when LLM evaluators are more likely to be unreliable and employ further human evaluation when necessary. To this end, we suggest combining the RTS and MCQ scores as a cost-efficient framework. Specifically, we calculate the correlation between RTS and MCQ scores for the  $i^{th}$  candidate system as a reliability indicator:
+
+$$
+\begin{array}{l} R _ {i} = \rho ([ f _ {\mathrm {R T S}} (g _ {i, 1}), \dots , f _ {\mathrm {R T S}} (g _ {i, N}) ], \tag {4} \\ \left[ f _ {\mathrm {M C Q}} \left(g _ {i, 1}\right), \dots , f _ {\mathrm {M C Q}} \left(g _ {i, N}\right) \right]) \\ \end{array}
+$$
+
+Then, we can loosely infer that up to a reliability tolerance  $r \in (0,1)$ , the LLM evaluators (either RTS or MCQ) are reliable if  $R_{i} > r$ . In other words, given a candidate  $i$ , if RTS and MCQ agree with each other up to a certain degree of tolerance  $r$ , we may assume the evaluator is reliable enough to avoid invoking further human evaluation.
+
+To validate this theory, we measure the correlations  $\rho (R_i,P_i^{\mathrm{RTS}})$  or  $\rho (R_i,P_i^{\mathrm{MCQ}})$ , where
+
+$P_{i}^{RTS / MCQ}$  is the performance of either method as defined in Equation (2). Given significantly large positive values of either  $\rho (R_i,P_i^{\mathrm{RTS}})$  or  $\rho (R_i,P_i^{\mathrm{MCQ}})$ , we can then conclude that  $R_{i}$  can be used as a reliable indicator for the performance of the corresponding method.
+
+As shown in Table 7,  $R_{i}$  demonstrates a significant correlation with  $P_{i}^{RTS}$  on both the consistency and fluency dimensions, and with  $P_{i}^{MCQ}$  on the coherence and consistency dimensions. This means that if RTS and MCQ generally agree with each other on the candidate's performance on a particular dimension with high  $\rho (R_{i},P_{i}^{\mathrm{RTS}})$  (or  $\rho (R_i,P_i^{\mathrm{MCQ}}))$ , RTS (or MCQ) is more likely to be human-aligned. Meanwhile, if RTS disagrees with MCQ  $(R_{i} < r)$ , further human evaluators are required to provide a conclusive evaluation. We provide  $R_{i}$  values for ChatGPT on each evaluated system in Appendix Table 29.
+
+# 6 Conclusion
+
+We explore the potential of using LLMs with different prompting techniques as metrics for abstractive summarization systems. Our extensive analysis suggests that while LLMs like ChatGPT perform better than commonly used automatic metrics across different summarization systems and dimensions, they are still not ready to replace human evaluators because they are candidate- and dimension-dependent, and they do not align well with human when comparing high-quality candidates. Nonetheless, if an LLM evaluator is to be used, we suggest combining multiple evaluation methods as a preliminary indicator to determine whether the metric is likely to be unreliable and whether further human evaluation is required.
+
+# Limitations
+
+Potential Human Bias. We benchmark the LLM evaluation results against the average of three human expert scores. Naturally, it is possible that these scores may exhibit potential biases of the human experts. Nevertheless, we wish to explore
+
+whether LLM evaluators are aligned with human experts, and may naturally exhibit the same bias as a human would. In other words, we examine whether we can reliably replace human annotators with LLMs, instead of seeking a "perfect" solution that has absolutely zero bias.
+
+Dataset Size. Given the constraints of the small size of the human-annotated SummEval dataset, we could only evaluate 100 summaries generated for each summarization system, with a total of 12 abstractive summarization systems. Since we have observed a significant correlation of LLM evaluations with humans for the consolidated 1200 summaries across all systems, it is possible that with a larger evaluation number, the per-system correlation could also be improved. In addition, given only 12 evaluated systems, our meta-correlation may still be subject to sample biases. We leave more investigations for the future once there are larger annotated datasets.
+
+Prompt tuning. Designing better prompt for LLMs are also ongoing research. Although it is possible that LLMs may act as better evaluators with better prompts, prompt tuning is not our focus. We seek to highlight the limitations of the investigated LLMs and have demonstrated that limitations such as negative meta-correlation are also found with a few other alternative prompts (see Appendix C).
+
+Availability of Commercialized LLM We note that the "gpt-3.5-turbo-0301" snapshot is currently taken down by OpenAI and replaced with a newer snapshot, "gpt-3.5-turbo-0613". This is also one disadvantage of using out-of-the-box commercialized LLM for summarization evaluations, as the exact checkpoints may not be stably available. As a result, future models may not be fairly compared against previously evaluated models using a different LLM checkpoint. Nevertheless, our paper only seeks to investigate the potential of LLM as an out-of-the-box evaluator, and the OpenAI models are currently one of the strongest. Eventually, we wish to raise awareness of some of the significant limitations found with these LLMs, which need to be resolved before LLMs can be used as direct replacements for human evaluations. In addition, we also note that the cost of evaluating only 100 sum
+
+maries for each system is relatively low (around 2 USD per system using ChatGPT). Since LLMs also conduct evaluations much faster than humans (around 2 minutes for LLMs versus 10 hours for human for 100 summaries), it may not pose significant barriers if one was to re-evaluate all compared systems on a single LLM.
+
+Limited Use of the Temporary Solution Unfortunately, our temporary efficient framework doesn't apply to the relevance dimension, where the  $R_{i}$  has no significant correlation with the performances of either RTS or MCQ. Moreover, the  $r$  value may be dataset-dependent, and it is hard to decide where to draw this line. We leave for future work of developing better methods to gauge the reliability of LLM evaluations.
+
+# Acknowledgements
+
+Yang You is being sponsored by NUS startup grant (Presidential Young Professorship), Singapore MOE Tier-1 grant, ByteDance grant, ARCTIC grant, SMI grant and Alibaba grant.
+
+# References
+
+Lidong Bing, Piji Li, Yi Liao, Wai Lam, Weiwei Guo, and Rebecca Passonneau. 2015. Abstractive multi-document summarization via phrase selection and merging. In Proceedings of ACL-IJCNLP.  
+Moye Chen, Wei Li, Jiachen Liu, Xinyan Xiao, Hua Wu, and Haifeng Wang. 2021. Sgsum: Transforming multi-document summarization into sub-graph selection. In Proceedings of EMNLP.  
+Liying Cheng, Xingxuan Li, and Lidong Bing. 2023. Is gpt-4 a good data analyst? arXiv preprint arXiv:2305.15038.  
+Liying Cheng, Dekun Wu, Lidong Bing, Yan Zhang, Zhanming Jie, Wei Lu, and Luo Si. 2020. Ent-description: Entity description generation by exploring knowledge graph. In Proceedings of EMNLP.  
+Cyril Chhun, Pierre Colombo, Fabian M. Suchanek, and Chloe Clavel. 2022. Of human criteria and automatic metrics: A benchmark of the evaluation of story generation. In Proceedings of Coling.  
+Cheng-Han Chiang and Hung-yi Lee. 2023. Can large language models be an alternative to human evaluations? In Proceedings of ACL.  
+Israel Cohen, Yiteng Huang, Jingdong Chen, Jacob Benesty, Jacob Benesty, Jingdong Chen, Yiteng Huang, and Israel Cohen. 2009. Pearson correlation coefficient. Noise reduction in speech processing.
+
+Yue Dong, Yikang Shen, Eric Crawford, Herke van Hoof, and Jackie Chi Kit Cheung. 2018. Banditsum: Extractive summarization as a contextual bandit. In Proceedings of EMNLP.  
+Ori Ernst, Avi Caciularu, Ori Shapira, Ramakanth Pasunuru, Mohit Bansal, Jacob Goldberger, and Ido Dagan. 2022. Proposition-level clustering for multi-document summarization. In Proceedings of NAACL.  
+Alexander Fabbri, Irene Li, Tianwei She, Suyi Li, and Dragomir Radev. 2019. Multi-news: A large-scale multi-document summarization dataset and abstractive hierarchical model. In Proceedings of ACL.  
+Alexander R Fabbri, Wojciech Krysciński, Bryan McCann, Caiming Xiong, Richard Socher, and Dragomir Radev. 2021. Summeval: Re-evaluating summarization evaluation. TACL.  
+Angela Fan, Mike Lewis, and Yann Dauphin. 2018. Hierarchical neural story generation. In Proceedings of ACL.  
+Luyu Gao, Aman Madaan, Shuyan Zhou, Uri Alon, Pengfei Liu, Yiming Yang, Jamie Callan, and Graham Neubig. 2022. Pal: Program-aided language models. In Proceedings of ICML.  
+Mingqi Gao, Jie Ruan, Renliang Sun, Xunjian Yin, Shipping Yang, and Xiaojun Wan. 2023. Human-like summarization evaluation with chatgpt. arXiv preprint arXiv:2304.02554.  
+Shen Gao, Xiuying Chen, Piji Li, Zhaochun Ren, Lidong Bing, Dongyan Zhao, and Rui Yan. 2019. Abstractive text summarization by incorporating reader comments. In Proceedings of AAAI.  
+Jian Guan and Minlie Huang. 2020. UNION: An Unreferenced Metric for Evaluating Open-ended Story Generation. In Proceedings of EMNLP.  
+Junxian He, Wojciech Kryscinski, Bryan McCann, Nazneen Rajani, and Caiming Xiong. 2022. CTRLsum: Towards generic controllable text summarization. In Proceedings of EMNLP.  
+Karl Moritz Hermann, Tomas Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching machines to read and comprehend. NeurIPS.  
+MG Kendall. 1938. A new measure of rank correlation. Biometrika.  
+Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. NeurIPS.  
+Wojciech Krysciński, Romain Paulus, Caiming Xiong, and Richard Socher. 2018. Improving abstraction in text summarization. In Proceedings of EMNLP.
+
+Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of ACL.  
+Piji Li, Wai Lam, Lidong Bing, and Zihao Wang. 2017. Deep recurrent generative decoder for abstractive text summarization. In Proceedings of EMNLP.  
+Wei Li, Xinyan Xiao, Jiachen Liu, Hua Wu, Haifeng Wang, and Junping Du. 2020. Leveraging graph to improve abstractive multi-document summarization. In Proceedings of ACL.  
+Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out.  
+Peter J Liu, Mohammad Saleh, Etienne Pot, Ben Goodrich, Ryan Sepassi, Lukasz Kaiser, and Noam Shazeer. 2018. Generating wikipedia by summarizing long sequences. In ICLR.  
+Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. 2023. G-eval: Nlg evaluation using gpt-4 with better human alignment. arXiv 2303.16634.  
+Yang Liu and Mirella Lapata. 2019. Hierarchical transformers for multi-document summarization. In Proceedings of ACL.  
+Yixin Liu, Ansong Ni, Linyong Nan, Budhaditya Deb, Chenguang Zhu, Ahmed H Awadallah, and Dragomir Radev. 2022. Leveraging locality in abstractive text summarization. In Proceedings of EMNLP.  
+Zheheng Luo, Qianqian Xie, and Sophia Ananiadou. 2023. Chatgpt as a factual inconsistency evaluator for abstractive text summarization. arXiv preprint arXiv:2303.15621.  
+Qing Lyu, Shreya Havaldar, Adam Stein, Li Zhang, Delip Rao, Eric Wong, Marianna Apidianaki, and Chris Callison-Burch. 2023. Faithful chain-of-thought reasoning. arXiv preprint arXiv:2301.13379.  
+Xueguang Ma, Xinyu Zhang, Ronak Pradeep, and Jimmy Lin. 2023a. Zero-shot listwise document reranking with a large language model. arXiv preprint arXiv:2305.02156.  
+Yubo Ma, Yixin Cao, YongChing Hong, and Aixin Sun. 2023b. Large language model is not a good few-shot information extractor, but a good reranker for hard samples! arXiv preprint arXiv:2303.08559.  
+Ani Nenkova. 2006. Summarization evaluation for text and speech: issues and approaches. In Ninth International Conference on Spoken Language Processing.  
+Ani Nenkova and Kathleen McKeown. 2012. A survey of text summarization techniques. Mining text data.  
+OpenAI. 2023. Gpt-4 technical report. arXiv preprint.
+
+Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. NeurIPS.  
+Karolina Owczarzak, John Conroy, Hoa Trang Dang, and Ani Nenkova. 2012. An assessment of the accuracy of automatic evaluation in summarization. In Proceedings of workshop on evaluation metrics and system comparison for automatic summarization.  
+Artidoro Pagnoni, Vidhisha Balachandran, and Yulia Tsvetkov. 2021. Understanding factuality in abstractive summarization with frank: A benchmark for factuality metrics. In Proceedings of NAACL-HLT.  
+Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of ACL.  
+Swarnadeep Saha, Shiyue Zhang, Peter Hase, and Mohit Bansal. 2022. Summarization programs: Interpretable abstractive summarization with neural modular trees. In *ICLR*.  
+Chenhui Shen, Liying Cheng, Lidong Bing, Yang You, and Luo Si. 2022a. Sentbs: Sentence-level beam search for controllable summarization. EMNLP.  
+Chenhui Shen, Liying Cheng, Xuan-Phi Nguyen, Lidong Bing, and Yang You. 2023. A hierarchical encoding-decoding scheme for abstractive multi-document summarization. In Findings of EMNLP.  
+Chenhui Shen, Liying Cheng, Ran Zhou, Lidong Bing, Yang You, and Luo Si. 2022b. Mred: A meta-review dataset for structure-controllable text generation. Findings of ACL.  
+C Spearman. 1987. The proof and measurement of association between two things. American Journal of Psychology.  
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288.  
+Jiaan Wang, Yunlong Liang, Fandong Meng, Haoxiang Shi, Zhixu Li, Jinan Xu, Jianfeng Qu, and Jie Zhou. 2023a. Is chatgpt a good nlg evaluator? a preliminary study. arXiv preprint arXiv:2303.04048.  
+Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, and Denny Zhou. 2023b. Self-consistency improves chain of thought reasoning in language models. In ICLR.  
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed H Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. NeurIPS.
+
+Johannes Welbl, Amelia Glaese, Jonathan Uesato, Sumanth Dathathri, John Mellor, Lisa Anne Hendricks, Kirsty Anderson, Pushmeet Kohli, Ben Coppin, and Po-Sen Huang. 2021. Challenges in detoxifying language models. In Findings of EMNLP.  
+Ning Wu, Ming Gong, Linjun Shou, Shining Liang, and Daxin Jiang. 2023. Large language models are diverse role-players for summarization evaluation. arXiv preprint arXiv:2303.15078.  
+Wen Xiao, Iz Beltagy, Giuseppe Carenini, and Arman Cohan. 2022. Primera: Pyramid-based masked sentence pre-training for multi-document summarization. In Proceedings of ACL.  
+Jiacheng Xu and Greg Durrett. 2021. Dissecting generation modes for abstractive summarization models via ablation and attribution. In Proceedings of ACLIJCNLP.  
+Kexin Yang, Dayiheng Liu, Wenqiang Lei, Baosong Yang, Mingfeng Xue, Boxing Chen, and Jun Xie. 2023. Tailor: A soft-prompt-based approach to attribute-based controlled text generation. In Proceedings of ACL.  
+Weizhe Yuan, Graham Neubig, and Pengfei Liu. 2021. Bartscore: Evaluating generated text as text generation. NeurIPS.  
+Jingqing Zhang, Yao Zhao, Mohammad Saleh, and Peter Liu. 2020a. Pegasus: Pre-training with extracted gap-sentences for abstractive summarization. In Proceedings of ICML.  
+Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2020b. Bertscore: Evaluating text generation with bert. In ICLR.  
+Qingyu Zhou, Nan Yang, Furu Wei, Shaohan Huang, Ming Zhou, and Tiejun Zhao. 2018. Neural document summarization by jointly learning to score and select sentences. In Proceedings of ACL.  
+Daniel M Ziegler, Nisan Stiannon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. 2019. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593.
+
+Score the following Summary given the corresponding Article with respect to consistency from one to five, where one indicates "inconsistency" and five indicates "perfect consistency". Note that consistency measures the factual alignment between the Summary and the Article, whether the Summary is faithful to the Article without introducing contradictions or misleading representations.
+
+Article: {article}
+
+Summary: {summary}
+
+Provide your reason in one sentence, then give a final score:
+
+Table 8: Example prompt for the RTS method on the consistency dimension. Text in {blue}: the specific article and the corresponding summary to be evaluated.
+
+Score the following Summary given the corresponding Article with respect to fluency from one to five, where one indicates "disfluency" and five indicates "perfect fluency". Note that fluency measures the quality of individual sentences in the Summary, whether the Summary is well-written, grammatically correct, and readable on the sentence level.
+
+Article: {article}
+
+Summary: {summary}
+
+Provide your reason in one sentence, then give a final score:
+
+Table 9: Example prompt for the RTS method on the fluency dimension. Text in {cyan}: the specific article and the corresponding summary to be evaluated.
+
+Score the following Summary given the corresponding Article with respect to coherence from one to five, where one indicates "incoherence" and five indicates "perfect coherence". Note that coherence measures the collective quality of the Summary, whether the Summary presents information that flows smoothly and avoids abrupt transitions or disjoint statements.
+
+Article: {article}
+
+Summary: {summary}
+
+Provide your reason in one sentence, then give a final score:
+
+Table 10: Example prompt for the RTS method on the coherence dimension. Text in {cyan}: the specific article and the corresponding summary to be evaluated.
+
+# A Evaluation Dimensions
+
+Fabbri et al. (2021) has defined 4 evaluation dimensions as follows:
+
+1. Coherence: The collective quality of all sentences. The summary should be well
+
+Choose an option from A to E in order to score the following Summary given the corresponding Article with respect to consistency from one to five, where one indicates "inconsistency" and five indicates "perfect consistency". Note that consistency measures the factual alignment between the Summary and the Article, whether the Summary is faithful to the Article without introducing contradictions or misleading representations.
+
+Article: {article}
+
+Summary: {summary}
+
+A: The Summary is totally inconsistent with the Article. Score: One.  
+B: The majority of the Summary is inconsistent with the Article. Score: Two.  
+C: Some information in the Summary is consistent with the Article whereas some are not. Score: Three.  
+D: The majority of the Summary is consistent with the Article. Score: Four.  
+E: All information included in the Summary is consistent with the Article. Score: Five.
+
+Your Answer (enter 1 letter from A to E):
+
+Table 11: Example prompt for the MCQ method on the consistency dimension. Text in {cyan}: the specific article and the corresponding summary to be evaluated.
+
+structured and well-organized. The summary should not just be a heap of related information but should build from sentence to sentence to a coherent body of information about a topic.
+
+2. Consistency: The factual alignment between the summary and the summarized source. A factually consistent summary contains only statements that are entailed by the source document.  
+3. Fluency: The quality of individual sentences. Sentences in the summary should have no formatting problems, capitalization errors or obviously ungrammatical sentences (e.g., fragments, missing components) that make the text difficult to read.  
+4. Relevance: Selection of important content from the source. The summary should include only important information from the source document.
+
+We follow the above definitions for designing ChatGPT's evaluation prompts.
+
+Choose an option from A to E in order to score the following Summary given the corresponding Article with respect to fluency from one to five, where one indicates "disfluency" and five indicates "perfect fluency". Note that fluency measures the quality of individual sentences in the Summary, whether the Summary is well-written, grammatically correct, and readable on the sentence level.
+
+Article: {article}
+
+Summary: {summary}
+
+A: The Summary is totally disfluent. Score: One.
+
+B: The majority of the Summary is disfluent. Score: Two.
+
+C: Some sentences in the Summary are fluent whereas some are not. Score: Three.
+
+D: The majority of the Summary is fluent. Score: Four
+
+E: All sentences in the Summary are fluent. Score: Five.
+
+Your Answer (enter 1 letter from A to E):
+
+Table 12: Example prompt for the MCQ method on the fluency dimension. Text in {cyan}: the specific article and the corresponding summary to be evaluated.
+
+Choose an option from A to E in order to score the following Summary given the corresponding Article with respect to coherence from one to five, where one indicates "incoherence" and five indicates "perfect coherence". Note that coherence measures the collective quality of the Summary, whether the Summary presents information that flows smoothly and avoids abrupt transitions or disjoint statements.
+
+Article: {article}
+
+Summary: {summary}
+
+A: The Summary is completely incoherent. Score: One.  
+B: The Summary is mostly incoherent. Score: Two.  
+C: The Summary is somewhat coherent. Score: Three.  
+D: The Summary is mostly coherent. Score: Four.  
+E: The Summary is completely coherent. Score: Five.
+
+Your Answer (enter 1 letter from A to E):
+
+Table 13: Example prompt for the MCQ method on the coherence dimension. Text in {cyan}: the specific article and the corresponding summary to be evaluated.
+
+# B Prompt Details and Design
+
+We show the RTS prompts for relevance, consistency, fluency, and coherence in Table 1, Table 8, Table 9, and Table 10 respectively.
+
+We show the MCQ prompts for relevance, consistency, fluency, and coherence in Table 2, Table 11, Table 12, and Table 13 respectively.
+
+We show the H2H prompts for relevance, consistency, fluency, and coherence in Table 3, Table 14, Table 15, and Table 16 respectively.
+
+Choose a more consistent summary from Summary #1 and Summary #2 with respect to the corresponding Article by choosing an option from A, B, or C. Note that consistency measures the factual alignment between the summary and the Article, whether the summary is faithful to the Article without introducing contradictions or misleading representations.
+
+Article: {article}
+
+Summary #1: {summary from model A}
+
+Summary #2: {summary from model B}
+
+A: Summary #1 is more consistent.
+
+B: Summary #2 is more consistent.
+
+C: Both Summary #1 and Summary #2 are equally consistent.
+
+Your choice (enter 1 letter from A to C):
+
+Table 14: Example prompt for the H2H method on the consistency dimension. Text in {cyan}: the specific article, and the corresponding summaries generated by a pair of compared models.
+
+Choose a more fluent summary from Summary #1 and Summary #2 with respect to the corresponding Article by choosing an option from A, B, or C. Note that fluency measures the quality of individual sentences in the summary, whether the summary is well-written, grammatically correct, and readable on the sentence level.
+
+Article: {article}
+
+Summary #1: {summary from model A}
+
+Summary #2: {summary from model B}
+
+A: Summary #1 is more fluent.
+
+B: Summary #2 is more fluent.
+
+C: Both Summary #1 and Summary #2 are equally fluent.
+
+Your choice (enter 1 letter from A to C):
+
+Table 15: Example prompt for the H2H method on the fluency dimension. Text in {cyan}: the specific article, and the corresponding summaries generated by a pair of compared models.
+
+To determine the exact definitions used in our prompts for each dimension, we re-use the first sentence from Fabbri et al. (2021)'s definition. We then prompt the LLM to provide a definition for the evaluated dimension, such as "define the word relevance in the context of summarization", then extract the key phrases generated that we believe to fit the definitions of Fabbri et al. (2021) to make up the full definition. We believe this approach may
+
+Choose a more coherent summary from Summary #1 and Summary #2 with respect to the corresponding Article by choosing an option from A, B, or C. Note that coherence measures the collective quality of the summary, whether the summary presents information that flows smoothly and avoids abrupt transitions or disjoint statements.
+
+Article: {article}
+
+Summary #1: {summary from model A}
+
+Summary #2: {summary from model B}
+
+A: Summary #1 is more coherent.
+
+B: Summary #2 is more coherent.
+
+C: Both Summary #1 and Summary #2 are equally coherent.
+
+Your choice (enter 1 letter from A to C):
+
+Table 16: Example prompt for the H2H method on the coherence dimension. Text in {cyan}: the specific article, and the corresponding summaries generated by a pair of compared models.  
+
+<table><tr><td></td><td>Coh</td><td>Con</td><td>Flu</td><td>Rel</td><td>Avg</td></tr><tr><td>ChatGPT-RTS2</td><td>5</td><td>6</td><td>7</td><td>6</td><td>6.00</td></tr><tr><td>ChatGPT-MCQ2</td><td>4</td><td>7</td><td>6</td><td>4</td><td>5.25</td></tr><tr><td>ChatGPT-StarEval</td><td>7</td><td>7</td><td>7</td><td>6</td><td>6.75</td></tr></table>
+
+Table 17: Total correct pairs for alternative prompts. Coh: coherence; Con: consistency; Flu: fluency; Rel: relevance.
+
+help the LLM to better evaluate the summaries according to definitions partially generated in its own language. Nevertheless, we didn't invest extensive efforts in prompt designs as this is not our key focus. We also demonstrate that our prompts have better evaluation results than two alternative prompts in Appendix C.
+
+# C Alternative prompts
+
+We also use ChatGPT to evaluate with the exact prompts from Wang et al. (2023a). We name these prompts "StarEval" since they prompt the LLM to give one to five stars for the summary. In addition, we use ChatGPT to evaluate with alternative prompts for RTS and MCQ by using the full definition as shown in Appendix A instead of supplementing the definition with ChatGPT-generated phrases. We name these two prompts RTS2 and MCQ2 respectively.
+
+We show the results of these alternative prompts in Table 17 and Table 18.
+
+# D Llama 2 Results
+
+We report the results of using three different sizes of Llama 2 models as LLM evaluators in Table 19. As shown, while the smallest model (7B) exhibits very low correlations with human scores (and only significant on the consistency and relevance dimensions), the larger models (13B and 70B) demonstrate significant correlations with human scores on the full dataset level. However, even the best-performing 70B model fails to outperform the human correlation of BARTScore, and is completely overwhelmed by the results of ChatGPT and GPT-4. This suggests that the open-sourced Llama 2 models are not suitable to be used as zero-shot evaluators. Moreover, all Llama 2 models exhibit significant meta-correlations for at least one dimension.
+
+<table><tr><td rowspan="2" colspan="2"></td><td colspan="3">Coherence</td><td colspan="3">Consistency</td><td colspan="3">Fluency</td><td colspan="3">Relevance</td></tr><tr><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td></tr><tr><td rowspan="2">RTS2</td><td>Human-Corr</td><td>0.339</td><td>0.338</td><td>0.285</td><td>0.393</td><td>0.497</td><td>0.350</td><td>0.290</td><td>0.280</td><td>0.252</td><td>0.440</td><td>0.441</td><td>0.351</td></tr><tr><td>Meta-Corr</td><td>-0.559</td><td>-0.476</td><td>-0.424</td><td>-0.748</td><td>-0.843</td><td>-0.576</td><td>-0.825</td><td>-0.823</td><td>-0.636</td><td>-0.385</td><td>-0.506</td><td>-0.212</td></tr><tr><td rowspan="2">MCQ2</td><td>Human-Corr</td><td>0.430</td><td>0.423</td><td>0.355</td><td>0.327</td><td>0.483</td><td>0.306</td><td>0.258</td><td>0.396</td><td>0.229</td><td>0.240</td><td>0.258</td><td>0.206</td></tr><tr><td>Meta-Corr</td><td>-0.308</td><td>-0.275</td><td>-0.212</td><td>-0.545</td><td>-0.446</td><td>-0.364</td><td>-0.811</td><td>-0.615</td><td>-0.636</td><td>-0.217</td><td>-0.707</td><td>-0.182</td></tr><tr><td rowspan="2">StarEval</td><td>Human-Corr</td><td>0.418</td><td>0.417</td><td>0.341</td><td>0.297</td><td>0.421</td><td>0.264</td><td>0.246</td><td>0.323</td><td>0.217</td><td>0.393</td><td>0.405</td><td>0.323</td></tr><tr><td>Meta-Corr</td><td>-0.084</td><td>-0.101</td><td>0.000</td><td>-0.664</td><td>-0.684</td><td>-0.545</td><td>-0.441</td><td>-0.460</td><td>-0.212</td><td>-0.497</td><td>-0.579</td><td>-0.303</td></tr></table>
+
+Table 18: Results of using alternative prompt with ChatGPT. Light gray values are insignificant (p-value  $\geq 0.05$ ). Human-Corr reports the overall correlation of ChatGPT scores with human scores. Meta-corr shows the meta-correlation.  
+
+<table><tr><td rowspan="2" colspan="2"></td><td colspan="3">Coherence</td><td colspan="3">Consistency</td><td colspan="3">Fluency</td><td colspan="3">Relevance</td></tr><tr><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td></tr><tr><td rowspan="2">7B</td><td>Human-Corr</td><td>-0.001</td><td>-0.000</td><td>-0.001</td><td>0.114</td><td>0.130</td><td>0.104</td><td>0.043</td><td>0.038</td><td>0.039</td><td>0.067</td><td>0.064</td><td>0.057</td></tr><tr><td>Meta-Corr</td><td>0.245</td><td>-0.258</td><td>0.182</td><td>-0.524</td><td>-0.720</td><td>-0.424</td><td>-0.042</td><td>0.463</td><td>-0.03</td><td>0.238</td><td>0.169</td><td>0.212</td></tr><tr><td rowspan="2">13B</td><td>Human-Corr</td><td>0.153</td><td>0.165</td><td>0.123</td><td>0.180</td><td>0.209</td><td>0.162</td><td>0.187</td><td>0.179</td><td>0.167</td><td>0.234</td><td>0.266</td><td>0.192</td></tr><tr><td>Meta-Corr</td><td>-0.287</td><td>-0.425</td><td>-0.182</td><td>-0.580</td><td>-0.495</td><td>-0.424</td><td>-0.455</td><td>-0.233</td><td>-0.303</td><td>-0.049</td><td>-0.406</td><td>0.000</td></tr><tr><td rowspan="2">70B</td><td>Human-Corr</td><td>0.234</td><td>0.254</td><td>0.186</td><td>0.357</td><td>0.395</td><td>0.319</td><td>0.155</td><td>0.161</td><td>0.134</td><td>0.248</td><td>0.285</td><td>0.200</td></tr><tr><td>Meta-Corr</td><td>-0.420</td><td>-0.407</td><td>-0.273</td><td>-0.811</td><td>-0.690</td><td>-0.667</td><td>-0.322</td><td>-0.319</td><td>-0.182</td><td>-0.238</td><td>-0.123</td><td>-0.182</td></tr></table>
+
+Table 19: Results of Llama 2 models of 7B, 13B, and 70B RTS correlations. Light gray values are insignificant (p-value  $\geq 0.05$ ). Human-Corr reports the overall correlation of LLM scores with human scores. Meta-corr shows the meta-correlation.  
+
+<table><tr><td>ID</td><td>Coh</td><td>Con</td><td>Flu</td><td>Rel</td></tr><tr><td>M8</td><td>0.58</td><td>0.09</td><td>0.13</td><td>0.51</td></tr><tr><td>M9</td><td>0.65</td><td>0.15</td><td>0.34</td><td>0.54</td></tr><tr><td>M10</td><td>0.58</td><td>0.22</td><td>0.20</td><td>0.54</td></tr><tr><td>M11</td><td>0.59</td><td>0.34</td><td>0.40</td><td>0.49</td></tr><tr><td>M12</td><td>0.65</td><td>0.03</td><td>0.14</td><td>0.54</td></tr><tr><td>M13</td><td>0.62</td><td>0.07</td><td>0.14</td><td>0.51</td></tr><tr><td>M14</td><td>0.65</td><td>0.09</td><td>0.18</td><td>0.54</td></tr><tr><td>M15</td><td>0.60</td><td>0.04</td><td>0.13</td><td>0.53</td></tr><tr><td>M17</td><td>0.58</td><td>0.05</td><td>0.08</td><td>0.52</td></tr><tr><td>M20</td><td>0.48</td><td>0.29</td><td>0.24</td><td>0.49</td></tr><tr><td>M22</td><td>0.58</td><td>0.02</td><td>0.14</td><td>0.54</td></tr><tr><td>M23</td><td>0.58</td><td>0.05</td><td>0.12</td><td>0.54</td></tr></table>
+
+# E Challenging Pairs
+
+To count the total correct pairs, we only evaluate the challenging pairs, which consist of summarization systems of consecutive performances according to average human scores across all dimensions. Thus, each pair contains 2 summarization systems with the smallest difference in terms of average performance.
+
+For instance, as shown in Table 21, M22 has the best average human score of 4.57, followed by M23 of 4.55, then M17 of 4.52. We thus compare model pairs of "M22-M23" and "M23-M17". The full challenge set is shown in Table 22.
+
+For RTS, MCQ, and all other baseline metrics,
+
+Table 20: The standard deviation of human annotations across different summarization systems and evaluation dimensions.  
+
+<table><tr><td>ID</td><td>Model Name</td><td>Coh</td><td>Con</td><td>Flu</td><td>Rel</td><td>Avg</td></tr><tr><td>M22</td><td>BART</td><td>4.18</td><td>4.94</td><td>4.90</td><td>4.25</td><td>4.57</td></tr><tr><td>M23</td><td>Pegasus (C4)</td><td>4.16</td><td>4.91</td><td>4.88</td><td>4.26</td><td>4.55</td></tr><tr><td>M17</td><td>T5</td><td>4.00</td><td>4.93</td><td>4.93</td><td>4.23</td><td>4.52</td></tr><tr><td>M12</td><td>Unified-ext-abs</td><td>3.60</td><td>4.96</td><td>4.85</td><td>3.85</td><td>4.32</td></tr><tr><td>M13</td><td>ROUGESal</td><td>3.44</td><td>4.82</td><td>4.86</td><td>3.83</td><td>4.24</td></tr><tr><td>M15</td><td>Closed book decoder</td><td>3.35</td><td>4.95</td><td>4.80</td><td>3.67</td><td>4.19</td></tr><tr><td>M14</td><td>Multi-task (Ent + QG)</td><td>3.20</td><td>4.90</td><td>4.74</td><td>3.63</td><td>4.12</td></tr><tr><td>M8</td><td>Pointer Generator</td><td>3.29</td><td>4.65</td><td>4.79</td><td>3.55</td><td>4.07</td></tr><tr><td>M9</td><td>Fast-abs-rl</td><td>2.38</td><td>4.67</td><td>4.50</td><td>3.52</td><td>3.77</td></tr><tr><td>M10</td><td>Bottom-Up</td><td>2.73</td><td>4.25</td><td>4.42</td><td>3.38</td><td>3.70</td></tr><tr><td>M20</td><td>GPT-2 (zero-shot)</td><td>3.63</td><td>3.40</td><td>3.97</td><td>3.30</td><td>3.58</td></tr><tr><td>M11</td><td>Improve-abs</td><td>2.28</td><td>3.27</td><td>3.65</td><td>3.15</td><td>3.09</td></tr></table>
+
+Table 21: The average human evaluation scores of various abstractive summarization models reported by Fabbri et al. (2021). We calculate the average (Avg) score of the reported coherence (Coh), consistency (Con), fluency (Flu), and relevance (Rel) scores. Rows are sorted according to the Avg column values in descending order.
+
+we simply need to compare the evaluated values across all systems, and each metric only needs to evaluate a total of 1200 summaries. However, for H2H, we need to evaluate a total of 6,600 summary pairs for the full standard set, and each pair needs to be evaluated twice with different summary positions (see § 3.1), resulting in a total of 13,200 LLM evaluations. Due to a limited budget, we thus only compare a challenge set of 11 pairs, reducing the total required LLM evaluations to 2,200.
+
+<table><tr><td rowspan="2">Model A</td><td rowspan="2">Model B</td><td colspan="3">Coherence</td><td colspan="3">Consistency</td><td colspan="3">Fluency</td><td colspan="3">Relevance</td></tr><tr><td>LLM</td><td>Human</td><td></td><td>LLM</td><td>Human</td><td></td><td>LLM</td><td>Human</td><td></td><td>LLM</td><td>Human</td><td></td></tr><tr><td>M22</td><td>M23</td><td>65.5</td><td>53.5</td><td>✓</td><td>55.75</td><td>52.5</td><td>✓</td><td>61</td><td>49.5</td><td>×</td><td>58.75</td><td>49.5</td><td>×</td></tr><tr><td>M23</td><td>M17</td><td>48.25</td><td>52.5</td><td>×</td><td>47</td><td>49</td><td>✓</td><td>49</td><td>45.5</td><td>✓</td><td>45</td><td>52</td><td>×</td></tr><tr><td>M17</td><td>M12</td><td>44</td><td>66.5</td><td>×</td><td>43.25</td><td>48.5</td><td>✓</td><td>40.5</td><td>54.5</td><td>×</td><td>49.25</td><td>72.5</td><td>×</td></tr><tr><td>M12</td><td>M13</td><td>58</td><td>51</td><td>✓</td><td>56.5</td><td>54.5</td><td>✓</td><td>56.75</td><td>50</td><td>×</td><td>58</td><td>45</td><td>×</td></tr><tr><td>M13</td><td>M15</td><td>45.5</td><td>49.5</td><td>✓</td><td>52</td><td>46</td><td>×</td><td>48.75</td><td>52</td><td>×</td><td>51.25</td><td>60.5</td><td>✓</td></tr><tr><td>M15</td><td>M14</td><td>57</td><td>54</td><td>✓</td><td>55</td><td>53.5</td><td>✓</td><td>56.5</td><td>52</td><td>✓</td><td>54</td><td>57</td><td>✓</td></tr><tr><td>M14</td><td>M8</td><td>49.25</td><td>44</td><td>✓</td><td>50</td><td>54.5</td><td>×</td><td>47</td><td>46.5</td><td>✓</td><td>49.25</td><td>53.5</td><td>×</td></tr><tr><td>M8</td><td>M9</td><td>77.5</td><td>82</td><td>✓</td><td>78.5</td><td>53</td><td>✓</td><td>80.5</td><td>63.5</td><td>✓</td><td>76</td><td>54</td><td>✓</td></tr><tr><td>M9</td><td>M10</td><td>45</td><td>36</td><td>✓</td><td>41.5</td><td>58</td><td>×</td><td>46</td><td>44.5</td><td>✓</td><td>41.5</td><td>56</td><td>×</td></tr><tr><td>M10</td><td>M20</td><td>58.25</td><td>24</td><td>×</td><td>61.75</td><td>64</td><td>✓</td><td>63.75</td><td>61.5</td><td>✓</td><td>61.5</td><td>54.5</td><td>✓</td></tr><tr><td>M20</td><td>M11</td><td>56.5</td><td>82</td><td>✓</td><td>50</td><td>53</td><td>×</td><td>51</td><td>58.5</td><td>✓</td><td>50</td><td>53</td><td>×</td></tr><tr><td colspan="2">#CP</td><td colspan="3">8</td><td colspan="3">7</td><td colspan="3">7</td><td colspan="3">4</td></tr></table>
+
+Table 22: #CP calculation for the ChatGPT-H2H metric of Model A over Model B. The numerical values in the middle section columns are aggregated scores for Model A. We omit the value for Model B, which is simply “100 - aggregated scores for Model A”. We use “√” to indicate both LLM and humans prefer the same model, and “×” otherwise. The model pairs are sorted in descending order according to the average human scores for each model.  
+
+<table><tr><td rowspan="2">ID</td><td colspan="3">Coherence</td><td colspan="3">Consistency</td><td colspan="3">Fluency</td><td colspan="3">Relevance</td></tr><tr><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td></tr><tr><td>M8</td><td>0.420</td><td>0.383</td><td>0.323</td><td>0.229</td><td>0.273</td><td>0.209</td><td>0.274</td><td>0.245</td><td>0.236</td><td>0.519</td><td>0.509</td><td>0.438</td></tr><tr><td>M9</td><td>0.174</td><td>0.243</td><td>0.142</td><td>0.119</td><td>0.265</td><td>0.103</td><td>0.256</td><td>0.258</td><td>0.219</td><td>0.254</td><td>0.307</td><td>0.200</td></tr><tr><td>M10</td><td>0.365</td><td>0.415</td><td>0.292</td><td>0.305</td><td>0.452</td><td>0.251</td><td>0.258</td><td>0.288</td><td>0.223</td><td>0.367</td><td>0.378</td><td>0.284</td></tr><tr><td>M11</td><td>0.378</td><td>0.420</td><td>0.320</td><td>0.404</td><td>0.488</td><td>0.335</td><td>0.227</td><td>0.288</td><td>0.182</td><td>0.501</td><td>0.511</td><td>0.394</td></tr><tr><td>M12</td><td>0.208</td><td>0.237</td><td>0.160</td><td>0.087</td><td>0.020</td><td>0.082</td><td>0.086</td><td>0.107</td><td>0.071</td><td>0.438</td><td>0.451</td><td>0.354</td></tr><tr><td>M13</td><td>0.455</td><td>0.473</td><td>0.359</td><td>0.178</td><td>0.037</td><td>0.167</td><td>0.063</td><td>0.151</td><td>0.055</td><td>0.403</td><td>0.403</td><td>0.329</td></tr><tr><td>M14</td><td>0.433</td><td>0.467</td><td>0.355</td><td>0.114</td><td>0.187</td><td>0.105</td><td>-0.007</td><td>0.055</td><td>-0.005</td><td>0.421</td><td>0.435</td><td>0.336</td></tr><tr><td>M15</td><td>0.372</td><td>0.385</td><td>0.279</td><td>0.189</td><td>0.316</td><td>0.177</td><td>0.100</td><td>0.087</td><td>0.085</td><td>0.252</td><td>0.288</td><td>0.193</td></tr><tr><td>M17</td><td>0.291</td><td>0.320</td><td>0.233</td><td>-0.086</td><td>-0.061</td><td>-0.084</td><td>0.017</td><td>0.046</td><td>0.015</td><td>0.204</td><td>0.199</td><td>0.175</td></tr><tr><td>M20</td><td>0.382</td><td>0.394</td><td>0.310</td><td>0.528</td><td>0.501</td><td>0.430</td><td>0.367</td><td>0.349</td><td>0.307</td><td>0.292</td><td>0.278</td><td>0.237</td></tr><tr><td>M22</td><td>0.215</td><td>0.293</td><td>0.162</td><td>-0.072</td><td>-0.052</td><td>-0.070</td><td>-0.114</td><td>-0.131</td><td>-0.101</td><td>0.201</td><td>0.300</td><td>0.172</td></tr><tr><td>M23</td><td>0.392</td><td>0.427</td><td>0.320</td><td>0.003</td><td>0.305</td><td>0.002</td><td>-0.078</td><td>-0.022</td><td>-0.069</td><td>0.223</td><td>0.324</td><td>0.184</td></tr></table>
+
+Table 23: Spearman (Spear.) correlations, Pearson (Pear.) correlations, and Kendall's Tau (Kend.) between ChatGPT-RTS and human scores on the 100 summaries for each model. Values in light gray color are insignificant  $(p\text{-value}\geq 0.05)$ .  
+
+<table><tr><td rowspan="2">ID</td><td colspan="3">Coherence</td><td colspan="3">Consistency</td><td colspan="3">Fluency</td><td colspan="3">Relevance</td></tr><tr><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td></tr><tr><td>M8</td><td>0.289</td><td>0.310</td><td>0.236</td><td>0.235</td><td>0.362</td><td>0.226</td><td>0.348</td><td>0.348</td><td>0.321</td><td>0.349</td><td>0.419</td><td>0.302</td></tr><tr><td>M9</td><td>0.170</td><td>0.170</td><td>0.148</td><td>0.211</td><td>0.321</td><td>0.200</td><td>0.198</td><td>0.299</td><td>0.174</td><td>0.318</td><td>0.356</td><td>0.276</td></tr><tr><td>M10</td><td>0.352</td><td>0.314</td><td>0.293</td><td>0.138</td><td>0.273</td><td>0.125</td><td>0.155</td><td>0.210</td><td>0.138</td><td>0.420</td><td>0.427</td><td>0.362</td></tr><tr><td>M11</td><td>0.285</td><td>0.312</td><td>0.250</td><td>0.380</td><td>0.370</td><td>0.317</td><td>0.200</td><td>0.218</td><td>0.163</td><td>0.397</td><td>0.428</td><td>0.334</td></tr><tr><td>M12</td><td>0.306</td><td>0.304</td><td>0.258</td><td>-0.025</td><td>-0.059</td><td>-0.024</td><td>0.256</td><td>0.306</td><td>0.239</td><td>0.283</td><td>0.287</td><td>0.246</td></tr><tr><td>M13</td><td>0.425</td><td>0.425</td><td>0.351</td><td>0.471</td><td>0.312</td><td>0.457</td><td>0.199</td><td>0.214</td><td>0.186</td><td>0.435</td><td>0.402</td><td>0.375</td></tr><tr><td>M14</td><td>0.490</td><td>0.472</td><td>0.422</td><td>0.112</td><td>0.157</td><td>0.110</td><td>0.084</td><td>0.143</td><td>0.078</td><td>0.326</td><td>0.329</td><td>0.284</td></tr><tr><td>M15</td><td>0.317</td><td>0.298</td><td>0.250</td><td>0.061</td><td>0.513</td><td>0.060</td><td>0.055</td><td>0.025</td><td>0.050</td><td>0.433</td><td>0.420</td><td>0.378</td></tr><tr><td>M17</td><td>0.250</td><td>0.255</td><td>0.215</td><td>-0.106</td><td>-0.081</td><td>-0.105</td><td>-0.011</td><td>0.024</td><td>-0.011</td><td>0.293</td><td>0.285</td><td>0.260</td></tr><tr><td>M20</td><td>0.463</td><td>0.450</td><td>0.381</td><td>0.455</td><td>0.442</td><td>0.371</td><td>0.494</td><td>0.450</td><td>0.404</td><td>0.326</td><td>0.334</td><td>0.264</td></tr><tr><td>M22</td><td>0.211</td><td>0.173</td><td>0.182</td><td>-0.096</td><td>-0.080</td><td>-0.095</td><td>-0.092</td><td>-0.056</td><td>-0.087</td><td>0.352</td><td>0.371</td><td>0.313</td></tr><tr><td>M23</td><td>0.218</td><td>0.209</td><td>0.189</td><td>0.107</td><td>0.448</td><td>0.104</td><td>-0.275</td><td>-0.269</td><td>-0.261</td><td>0.147</td><td>0.187</td><td>0.129</td></tr></table>
+
+Table 24: Spearman (Spear.) correlations, Pearson (Pear.) correlations, and Kendall's Tau (Kend.) between ChatGPT-MCQ and human scores on the 100 summaries for each model. Values in light gray color are insignificant (p-value  $\geq 0.05$ ).
+
+# F Average ChatGPT scores
+
+We present the average ChatGPT evaluation scores for each model across all dimensions in Table 25. Generally, the same trend holds for the individual systems, that ChatGPT score systems much more
+
+conservatively with RTS, and becomes more optimistic with MCQ.
+
+<table><tr><td rowspan="2">ID</td><td colspan="4">Coherence</td><td colspan="4">Consistency</td><td colspan="4">Fluency</td><td colspan="4">Relevance</td></tr><tr><td>Chat-RTS</td><td>Chat-MCQ</td><td>GPT-4</td><td>human</td><td>Chat-RTS</td><td>Chat-MCQ</td><td>GPT-4</td><td>human</td><td>Chat-RTS</td><td>Chat-MCQ</td><td>GPT-4</td><td>human</td><td>Chat-RTS</td><td>Chat-MCQ</td><td>GPT-4</td><td>human</td></tr><tr><td>M8</td><td>2.43</td><td>3.70</td><td>4.56</td><td>3.29</td><td>4.21</td><td>4.70</td><td>4.77</td><td>4.65</td><td>2.51</td><td>3.76</td><td>4.61</td><td>4.79</td><td>3.56</td><td>4.33</td><td>4.71</td><td>3.55</td></tr><tr><td>M9</td><td>1.80</td><td>3.51</td><td>4.41</td><td>2.38</td><td>3.93</td><td>4.69</td><td>4.94</td><td>4.67</td><td>2.16</td><td>3.45</td><td>4.15</td><td>4.50</td><td>3.51</td><td>4.33</td><td>4.82</td><td>3.52</td></tr><tr><td>M10</td><td>2.05</td><td>3.49</td><td>4.05</td><td>2.73</td><td>3.77</td><td>4.52</td><td>4.60</td><td>4.25</td><td>2.23</td><td>3.43</td><td>3.92</td><td>4.42</td><td>3.45</td><td>4.25</td><td>4.62</td><td>3.38</td></tr><tr><td>M11</td><td>1.70</td><td>2.63</td><td>2.93</td><td>2.28</td><td>2.36</td><td>4.11</td><td>3.72</td><td>3.27</td><td>1.71</td><td>2.66</td><td>2.77</td><td>3.65</td><td>2.92</td><td>4.06</td><td>3.87</td><td>3.15</td></tr><tr><td>M12</td><td>2.75</td><td>3.92</td><td>4.75</td><td>3.60</td><td>4.46</td><td>4.75</td><td>5.00</td><td>4.96</td><td>2.59</td><td>3.84</td><td>4.72</td><td>4.85</td><td>3.89</td><td>4.40</td><td>4.95</td><td>3.85</td></tr><tr><td>M13</td><td>2.91</td><td>3.99</td><td>4.76</td><td>3.44</td><td>4.40</td><td>4.73</td><td>4.89</td><td>4.82</td><td>2.69</td><td>3.93</td><td>4.69</td><td>4.86</td><td>3.90</td><td>4.41</td><td>4.85</td><td>3.83</td></tr><tr><td>M14</td><td>2.38</td><td>3.87</td><td>4.54</td><td>3.20</td><td>4.25</td><td>4.70</td><td>4.99</td><td>4.90</td><td>2.42</td><td>3.83</td><td>4.53</td><td>4.74</td><td>3.67</td><td>4.32</td><td>4.84</td><td>3.63</td></tr><tr><td>M15</td><td>2.65</td><td>3.81</td><td>4.58</td><td>3.35</td><td>4.33</td><td>4.75</td><td>4.90</td><td>4.95</td><td>2.60</td><td>3.84</td><td>4.61</td><td>4.80</td><td>3.78</td><td>4.32</td><td>4.82</td><td>3.67</td></tr><tr><td>M17</td><td>3.05</td><td>4.11</td><td>4.89</td><td>4.00</td><td>4.84</td><td>4.87</td><td>4.97</td><td>4.93</td><td>3.29</td><td>4.08</td><td>4.79</td><td>4.93</td><td>4.30</td><td>4.43</td><td>4.97</td><td>4.23</td></tr><tr><td>M20</td><td>2.00</td><td>2.95</td><td>2.99</td><td>3.63</td><td>2.66</td><td>3.57</td><td>3.51</td><td>3.40</td><td>2.22</td><td>3.02</td><td>3.05</td><td>3.97</td><td>2.73</td><td>3.78</td><td>3.47</td><td>3.30</td></tr><tr><td>M22</td><td>3.54</td><td>4.09</td><td>4.92</td><td>4.18</td><td>4.83</td><td>4.82</td><td>4.97</td><td>4.94</td><td>3.57</td><td>4.06</td><td>4.94</td><td>4.90</td><td>4.41</td><td>4.43</td><td>4.95</td><td>4.25</td></tr><tr><td>M23</td><td>3.00</td><td>4.14</td><td>4.88</td><td>4.16</td><td>4.74</td><td>4.84</td><td>4.94</td><td>4.91</td><td>3.08</td><td>4.05</td><td>4.76</td><td>4.88</td><td>4.10</td><td>4.45</td><td>4.96</td><td>4.26</td></tr><tr><td>avg</td><td>2.52</td><td>3.68</td><td>4.35</td><td>3.35</td><td>4.06</td><td>4.59</td><td>4.68</td><td>4.55</td><td>2.59</td><td>3.66</td><td>4.29</td><td>4.61</td><td>3.68</td><td>4.29</td><td>4.65</td><td>3.72</td></tr></table>
+
+Table 25: Average scores for ChatGPT using RTS (Chat-RTS) and MCQ (Chat-MCQ), GPT-4 using RTS (GPT-4), as well as the human scores.  
+Dimension Generation  
+
+<table><tr><td>Consistency</td><td>The summary is inconsistent with the article as it omits important details and repeats a phrase multiple times, resulting in a score of 1.</td></tr><tr><td>Consistency</td><td>The summary includes some relevant information but lacks coherence and omits important details, resulting in a score of 2.5 out of 5 for consistency.</td></tr><tr><td>Fluency</td><td>The summary is well-written and grammatically correct, but lacks important details from the article such as the potential link between antibiotics and weight gain, and the warnings about over-prescription of antibiotics and the development of superbugs. Score: 3/5.</td></tr><tr><td>Fluency</td><td>The summary lacks coherence and does not provide a clear understanding of the article, resulting in a score of 1 for disfluency.</td></tr></table>
+
+Table 26: Examples of wrong reasons generated during RTS by ChatGPT that do not correspond to the evaluated dimension. Bolded: reasons that don't match the evaluated dimension.
+
+# G Incorrect Reasons in RTS
+
+We illustrate some of the observed incorrect reasons generated with the RTS method by ChatGPT that do not correspond to the evaluated dimension in Table 26. We further provide incorrect reasons generated by GPT-4 in Table 27.
+
+We haven't compiled the exact number of mismatched reasons for either LLM, but we do observe a non-trivial fraction of the evaluations containing unrelated-to-dimension reasoning during preliminary verification. For instance, on the consistency dimension, ChatGPT-RTS has  $42\%$  evaluations containing dimension-irrelevant reasoning for the M11 (avg. score of 2.36) model, and  $10\%$  for the M17 model (avg. score of 4.84).
+
+Dimension Generation  
+
+<table><tr><td>Coherence</td><td>The summary is somewhat coherent but includes incorrect information, such as stating that Merson restarted his row with Burnley instead of Andros Townsend, and does not flow smoothly; therefore, I would give it a score of 3.</td></tr><tr><td>Consistency</td><td>The summary is consistent with the article as it mentions the helicopter airlifting patients to the hospital, but it leaves out important details such as the number of people involved and their injuries. Final score: 3.</td></tr><tr><td>Consistency</td><td>The summary accurately reflects the main points of the article, although it could have mentioned Falcao&#x27;s unhappiness in France. Final score: 4.</td></tr><tr><td>Fluency</td><td>The summary is coherent and grammatically correct, but it does not provide enough information about the incident. Final score: 3.</td></tr><tr><td>Fluency</td><td>The summary is only one sentence and do not provide enough information about the article&#x27;s content. Final score: 1.</td></tr></table>
+
+Table 27: Examples of wrong reasons generated during RTS by GPT-4 that don't correspond to the evaluated dimension. Bolded: reasons that don't match the evaluated dimension.
+
+# H GPT-4 Evaluator
+
+We also look into the reasoning of GPT-4 and discover that it makes the same mistakes as ChatGPT by penalizing the summary for reasons unrelated to the evaluated dimension (see Table 27).
+
+Another major difference is that GPT-4 tends to give overly generous scores. In one exceptionally extreme case, GPT-4 gives full scores for all generations by M12 in terms of consistency. Table 25 also shows the much higher average scores given by GPT-4 across all dimensions than those of ChatGPT-RTS.
+
+<table><tr><td rowspan="2">ID</td><td colspan="3">Coherence</td><td colspan="3">Consistency</td><td colspan="3">Fluency</td><td colspan="3">Relevance</td></tr><tr><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td></tr><tr><td>M8</td><td>0.429</td><td>0.446</td><td>0.362</td><td>0.449</td><td>0.597</td><td>0.433</td><td>0.412</td><td>0.409</td><td>0.385</td><td>0.425</td><td>0.489</td><td>0.365</td></tr><tr><td>M9</td><td>0.288</td><td>0.346</td><td>0.243</td><td>0.349</td><td>0.350</td><td>0.331</td><td>0.462</td><td>0.465</td><td>0.407</td><td>0.413</td><td>0.490</td><td>0.358</td></tr><tr><td>M10</td><td>0.495</td><td>0.480</td><td>0.403</td><td>0.518</td><td>0.683</td><td>0.469</td><td>0.406</td><td>0.579</td><td>0.360</td><td>0.510</td><td>0.537</td><td>0.429</td></tr><tr><td>M11</td><td>0.584</td><td>0.572</td><td>0.474</td><td>0.529</td><td>0.536</td><td>0.439</td><td>0.448</td><td>0.436</td><td>0.358</td><td>0.500</td><td>0.519</td><td>0.405</td></tr><tr><td>M12</td><td>0.245</td><td>0.360</td><td>0.209</td><td>-</td><td>-</td><td>-</td><td>0.387</td><td>0.506</td><td>0.372</td><td>0.169</td><td>0.106</td><td>0.149</td></tr><tr><td>M13</td><td>0.271</td><td>0.271</td><td>0.230</td><td>0.535</td><td>0.444</td><td>0.518</td><td>0.093</td><td>0.337</td><td>0.089</td><td>0.237</td><td>0.337</td><td>0.206</td></tr><tr><td>M14</td><td>0.263</td><td>0.312</td><td>0.222</td><td>-0.040</td><td>-0.026</td><td>-0.040</td><td>0.311</td><td>0.432</td><td>0.291</td><td>0.277</td><td>0.315</td><td>0.241</td></tr><tr><td>M15</td><td>0.403</td><td>0.418</td><td>0.342</td><td>0.158</td><td>0.504</td><td>0.155</td><td>0.240</td><td>0.242</td><td>0.225</td><td>0.307</td><td>0.402</td><td>0.267</td></tr><tr><td>M17</td><td>0.285</td><td>0.240</td><td>0.246</td><td>0.390</td><td>0.748</td><td>0.385</td><td>0.091</td><td>0.131</td><td>0.088</td><td>0.210</td><td>0.253</td><td>0.185</td></tr><tr><td>M20</td><td>0.427</td><td>0.418</td><td>0.334</td><td>0.378</td><td>0.362</td><td>0.313</td><td>0.482</td><td>0.468</td><td>0.402</td><td>0.479</td><td>0.491</td><td>0.391</td></tr><tr><td>M22</td><td>0.026</td><td>0.004</td><td>0.023</td><td>0.346</td><td>0.608</td><td>0.343</td><td>0.248</td><td>0.141</td><td>0.242</td><td>0.143</td><td>0.084</td><td>0.126</td></tr><tr><td>M23</td><td>0.320</td><td>0.330</td><td>0.283</td><td>0.360</td><td>0.168</td><td>0.354</td><td>0.267</td><td>0.136</td><td>0.251</td><td>0.005</td><td>0.003</td><td>0.005</td></tr></table>
+
+Table 28: Spearman (Spear.) correlations, Pearson (Pear.) correlations, and Kendall's Tau (Kend.) between GPT-4 RTS and human scores on the 100 summaries for each model. Values in light gray color are insignificant (p-value  $\geq 0.05$ ). Note that for the consistency of M12, correlations cannot be calculated because GPT-4 gives 5 scores to all examples.  
+
+<table><tr><td rowspan="2">ID</td><td colspan="3">Coherence</td><td colspan="3">Consistency</td><td colspan="3">Fluency</td><td colspan="3">Relevance</td></tr><tr><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td><td>Spear.</td><td>Pear.</td><td>Kend.</td></tr><tr><td>M8</td><td>0.464</td><td>0.455</td><td>0.407</td><td>0.557</td><td>0.613</td><td>0.526</td><td>0.391</td><td>0.367</td><td>0.341</td><td>0.517</td><td>0.514</td><td>0.471</td></tr><tr><td>M9</td><td>0.436</td><td>0.415</td><td>0.416</td><td>0.215</td><td>0.342</td><td>0.195</td><td>0.578</td><td>0.571</td><td>0.517</td><td>0.450</td><td>0.485</td><td>0.412</td></tr><tr><td>M10</td><td>0.374</td><td>0.345</td><td>0.340</td><td>0.300</td><td>0.297</td><td>0.270</td><td>0.631</td><td>0.585</td><td>0.551</td><td>0.506</td><td>0.507</td><td>0.459</td></tr><tr><td>M11</td><td>0.326</td><td>0.428</td><td>0.307</td><td>0.437</td><td>0.421</td><td>0.393</td><td>0.354</td><td>0.509</td><td>0.303</td><td>0.467</td><td>0.486</td><td>0.419</td></tr><tr><td>M12</td><td>0.500</td><td>0.469</td><td>0.444</td><td>0.446</td><td>0.486</td><td>0.424</td><td>0.273</td><td>0.320</td><td>0.241</td><td>0.588</td><td>0.591</td><td>0.542</td></tr><tr><td>M13</td><td>0.462</td><td>0.450</td><td>0.408</td><td>0.470</td><td>0.510</td><td>0.445</td><td>0.325</td><td>0.345</td><td>0.282</td><td>0.515</td><td>0.498</td><td>0.475</td></tr><tr><td>M14</td><td>0.524</td><td>0.460</td><td>0.463</td><td>0.325</td><td>0.293</td><td>0.307</td><td>0.216</td><td>0.257</td><td>0.194</td><td>0.501</td><td>0.476</td><td>0.455</td></tr><tr><td>M15</td><td>0.534</td><td>0.499</td><td>0.470</td><td>0.266</td><td>0.326</td><td>0.251</td><td>0.383</td><td>0.369</td><td>0.337</td><td>0.584</td><td>0.550</td><td>0.535</td></tr><tr><td>M17</td><td>0.312</td><td>0.294</td><td>0.274</td><td>-0.016</td><td>-0.008</td><td>-0.015</td><td>0.251</td><td>0.261</td><td>0.225</td><td>0.530</td><td>0.479</td><td>0.504</td></tr><tr><td>M20</td><td>0.636</td><td>0.623</td><td>0.566</td><td>0.733</td><td>0.664</td><td>0.644</td><td>0.566</td><td>0.556</td><td>0.492</td><td>0.570</td><td>0.565</td><td>0.496</td></tr><tr><td>M22</td><td>0.314</td><td>0.338</td><td>0.276</td><td>0.245</td><td>0.151</td><td>0.241</td><td>0.371</td><td>0.341</td><td>0.330</td><td>0.471</td><td>0.412</td><td>0.451</td></tr><tr><td>M23</td><td>0.380</td><td>0.386</td><td>0.337</td><td>0.158</td><td>0.496</td><td>0.153</td><td>0.296</td><td>0.275</td><td>0.259</td><td>0.450</td><td>0.505</td><td>0.421</td></tr></table>
+
+Table 29:  $R_{i}$ , the reliability indicator calculated by the Spearman (Spear.) correlations, Pearson (Pear.) correlations, and Kendall's Tau (Kend.) between ChatGPT-RTS and ChatGPT-MCQ. Values in light gray color are insignificant (p-value ≥ 0.05).

2023/Large Language Models are Not Yet Human-Level Evaluators for Abstractive Summarization/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:105a8c3fbaac65470be897e02e8e19622d0525b090a78f444146e7f622382faf
+size 1180902

2023/Large Language Models as Source Planner for Personalized Knowledge-grounded Dialogues/3db068d7-7569-4325-831f-81a84ff080cc_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e4f3cecbcd853d6ce9d653dd7e46d30654704f8ce0d70eb098c520d286354fd9
+size 1008545

2023/Large Language Models as Source Planner for Personalized Knowledge-grounded Dialogues/full.md

@@ -0,0 +1,419 @@
+# Large Language Models as Source Planner for Personalized Knowledge-grounded Dialogues
+
+Hongru Wang, Minda Hu, Yang Deng*, Rui Wang, Fei Mi, Weichao Wang, Yasheng Wang, Wai-Chung Kwan, Irwin King, Kam-Fai Wong
+
+$\diamond$ Department of Systems Engineering and Engineering Management,
+
+$\clubsuit$ Department of Computer Science and Engineering, The Chinese University of Hong Kong
+
+$\spadesuit$ National University of Singapore,  $\diamond$ Huawei Noah's Ark Lab {hrwang, kfwong}@se.cuhk.edu.hk ydeng@nus.edu.sg
+
+# Abstract
+
+Open-domain dialogue system usually requires different sources of knowledge to generate more informative and evidential responses. However, existing knowledge-grounded dialogue systems either focus on a single knowledge source or overlook the dependency between multiple sources of knowledge, which may result in generating inconsistent or even paradoxical responses. To incorporate multiple knowledge sources and dependencies between them, we propose SAFARI, a novel framework that leverages the exceptional capabilities of large language models (LLMs) in planning, understanding, and incorporating under both supervised and unsupervised settings. Specifically, SAFARI decouples the knowledge grounding into multiple knowledge sources selection and response generation, which allows easy extension to various knowledge sources including the possibility of not using any sources. To study the problem, we construct a personalized knowledge-grounded dialogue dataset Knowledge Behind Persona (KBP), which is the first to consider the dependency between persona and implicit knowledge. Experimental results on the KBP dataset demonstrate that the SAFARI framework can effectively produce persona-consistent and knowledge-enhanced responses.
+
+# 1 Introduction
+
+Knowledge enhancement techniques (Yu et al., 2022) have significantly empowered machines to deepen their understanding of the underlying knowledge in open-domain dialogues (Huang et al., 2020), surpassing what can be solely acquired from conversational corpora. Recent years have witnessed various open-domain dialogue systems relying on different types of knowledge sources, such as topic (Wang et al., 2022; Zhu et al., 2023), external documents (e.g., Wikipedia) (Dinan et al.,
+
+2019; Zhou et al., 2020; Wang et al., 2023c; Deng et al., 2023c), persona (Zhang et al., 2018; Liu et al., 2022a; Wang et al., 2023b; Deng et al., 2023b), user memory (Xu et al., 2022c,b; Park et al., 2023; Deng et al., 2022), and more. Realizing the limitations of using single-source knowledge, some recent studies further develop dialogue systems with access to multi-source knowledge (Wu et al., 2021, 2022b; Jang et al., 2022).
+
+Despite their effectiveness of enriching the dialogue responses with multi-source knowledge, existing methods typically design models to incorporate all sources indiscriminately, resulting in a cumbersome process that struggles to handle cases dependent on the interaction between some specific sources instead of all (Wu et al., 2022b; Fu et al., 2022). Moreover, the importance of comprehending the potential dependency between knowledge sources is overlooked in previous works, which may result in generating paradoxical responses (Majumder et al., 2020; Jang et al., 2022). For example, humans often express their persona with the assistance of external knowledge. As shown in Figure 1(a), for responding to the question "Hi, what do you like to eat?", it is inadequate to only incorporate single-source knowledge from user persona, e.g., "I am a vegetarian", since relevant information from external documents is also required, e.g., Vegetarian (Vegetarians like to eat fruits and vegetables). However, being unaware of the dependency between these two different sources of knowledge (persona and documents), dialogue systems may select the document implying inconsistent personas (e.g., "Food contains meat, fruits, ..."), leading to responses conflicting with defined personas (e.g., "I like to eat meat"). Therefore, it attaches great importance in modeling the interaction and dependency of different sources<sup>1</sup> in building knowledge-grounded dialogue systems.
+
+The absence of problem definitions and well
+
+![](images/236f5813d5e13801dfdde72e5bc362b9231be1918c5189c117d81e24a43c5c69.jpg)  
+Figure 1: (a) An example of dependency of two sources involved in the persona-consistent dialogue system (PERSONA and DOCUMENTS); (b) our proposed SAFARI framework to plan, retrieve, and incorporate multiple sources of knowledge: PERSONA, DOCUMENTS, and so on. Planning, Retrieval and Assembling steps are divided by dashed lines; (c) A sample from the KBP dataset. There are three situations of responses in our datasets: 1) response without the need for any sources (NULL), 2) response using only personae description (from PERSONA source), and 3) response using both persona and knowledge (from PERSONA, DOCUMENTS sources). The example here presents the first and third situations. We highlight the response and used knowledge with the same color.
+
+![](images/7b275051b4e359249fb003eb0ba56e399b8f79438bb8092899492afb195ada82.jpg)
+
+![](images/6d5bc24df615d3641f87a134f7a110573c23950c9fe718b170093c63c75b4bbe.jpg)
+
+established benchmarks greatly impedes the progress of building dialogue systems that can capture the knowledge dependency between different sources. To this end, we construct a personalized knowledge-grounded dialogue dataset, named Knowledge Behind Persona (KBP), to mimic the scenario where the understanding of persona-knowledge dependency is required to produce consistent responses. KBP aims to build better persona-consistent dialogue systems with the help of utilizing underlying knowledge behind different persona descriptions comprehensively.
+
+In order to address the interaction and dependency issue between specific sources, we propose a novel framework, named Source plAnner For personAlized knowledge-gRounded dDialogues (SAFARI). Seeing the potential of large language models (LLMs) in planning the use of external information (Schick et al., 2023; Shen et al., 2023; Gao et al., 2023), we explore LLMs' capability of connecting different sources in the context of the personalized knowledge-grounded dialogue system. As illustrated in Figure 1(b), the whole response generation can be modeled into three steps in SAFARI: 1) Planning to make a series of decisions of whether to use a specific knowledge source given relationship descriptions between different sources; 2) Retrieval to retrieve top-n results from external databases according to the decisions; 3) Assembling to incorporate all retrieved knowledge into the final response generation. Benefiting from decoupling source selection and response generation, our framework is more flexible and scalable, allowing independent modification of each component.
+
+Additionally, our framework can easily accommodate scenarios where multiple or no sources are required. To sum up, our contributions are listed below:
+
+- We propose the SAFARI framework to augment the dialogue system to plan and incorporate multiple sources of knowledge into responses ((e.g., decide whether or not require knowledge, which source to call, and when to call)), and further address the knowledge dependency issue between sources in both supervised and unsupervised manners by leveraging LLMs.  
+- We build a personalized knowledge-grounded dialogue dataset, KBP, where the responses are conditioned on multiple sources of knowledge, leading to more user-engaged dialogues with informative and persona-consistent knowledge.  
+- We conduct exhaustive experiments to validate the effectiveness of our proposed framework to incorporate multiple sources and capture the dependency between them<sup>2</sup>.
+
+# 2 Related Works
+
+# 2.1 Personalized Dialogue System
+
+To build a personalized dialog agent, Zhang et al. (2018) extensively investigated this task with a new dataset Persona-Chat, where a pre-defined persona set is a form of multiple sentences of textual description. Lots of works follow this setting and have taken mutual persona perception (Liu
+
+et al., 2020; Kim et al., 2020; Xu et al., 2022a), persona-sparse scenario (Song et al., 2021; Welch et al., 2022), long-term persona memory (Xu et al., 2022c), persona extending (Liu et al., 2022b) and persona order bias (Chen et al., 2023) into consideration. Although some of them complement the insufficient semantics in short persona descriptions by further utilizing an external commonsense knowledge base to extend existing persona sets (Majumder et al., 2020; Liu et al., 2022b), they still fall into the conventional framework coupling the knowledge selection with the response generation (Wu et al., 2022b), rendering it infeasible to handle various sources of knowledge. There have also been works showing that the combination of different knowledge sources such as persona descriptions and Wikipedia can further improve the overall performance (Jang et al., 2022; Wu et al., 2021, 2022a). However, they still fail to capture possible dependency between knowledge sources. In their framework, knowledge is not used as the role to assist persona-consistent response generation, but as an additional resource to generate a more informative response (Dinan et al., 2019; Xue et al., 2023) or select a suitable persona (Jang et al., 2022; Fu et al., 2022).
+
+# 2.2 LLMs for Planning
+
+Large Language Models (LLMs) show remarkable capabilities in planning the use of various external resources, such as tools (Schick et al., 2023), models (Shen et al., 2023), and APIs (Li et al., 2023), to solve various NLP tasks and suit different applications in practice. Alternatively, different types of knowledge can be retrieved from external sources, as illustrated in WebGPT (Nakano et al., 2022) and ReAct (Yao et al., 2023). Integrating various knowledge sources to improve the quality of LLM generation becomes increasingly challenging due to the need for strategic planning, sequential decision-making, and complex reasoning. Previous research primarily focuses on either earlier decision-making stages (Nakano et al., 2022; Shen et al., 2023) or the subsequent response generation (Sun et al., 2023; Schick et al., 2023; Deng et al., 2023a), instead of establishing a complete framework for planning the use of multiple knowledge sources to generate appropriate responses. There is a latest work named TPE which regards different knowledge sources as conceptual tools and proposes a multi-persona collaboration framework to
+
+model the decision-making process of the call order for multiple knowledge sources (Wang et al., 2023a). We differ in exploring the planning capability of LLMs to decide whether or not require knowledge, which source to call, and when to call in both the supervised and unsupervised manner.
+
+# 3 Data Collection
+
+In this section, we dually introduce the process of data collection and statistics of the collected data. The data collection process can be divided into two steps: Step 1. Persona and Knowledge Acquisition and Step 2. Dialog Collection.
+
+# 3.1 Persona and Knowledge Acquisition
+
+Seeds Preparation. To reduce annotation cost, we take advantage of the currently available persona dialogue dataset: DuLemon (Xu et al., 2022c) and two widely-adopted Chinese knowledge bases: Baike<sup>3</sup> and Ownthink<sup>4</sup> to produce seed data. Specifically, we first cluster all persona sentences from DuLeMon into 10 topics. After removing duplicate, rare, and similar personas, we carefully choose around 20 personas for each left topic as seed personas<sup>5</sup>. In addition, we manually add some personas for the existing topics and new topics. The final personas consist of age, nation, personality, career, movie, music, sport, book, constellation, locality, gender, others. For retrieving persona-related knowledge, we simply combine two aforementioned knowledge bases with similar filtering operations and store them as (head entity, attribute, tail entity) tuples.
+
+Persona and Knowledge Matching. For each persona sentence, we segment it into a sequence of words with a Chinese word segmentation tool jieba<sup>6</sup>. If any words exactly match the head entity or tail entity of a certain knowledge tuple, we transform the tuple into a sentence according to predefined templates and then save it as one knowledge for this persona sentence. In this way, we can obtain various knowledge for each persona sentence. Consistent with previous works (Zhang et al., 2018; Majumder et al., 2020), we randomly sample 3 persona sentences along with 5 knowledge sentences per persona to form a persona description
+
+<sup>3</sup>http://www.openkg.cn  
+<sup>4</sup>http://github.com/ownthink/KnowledgeGraphData  
+Some topics are removed if they contain less than 20 personas. We don't pick hundreds of personas because one persona sentence has a vast amount of knowledge behind it.  
+$^{6}$ https://github.com/fxsjy/jieba
+
+of the system for each dialog. More details and templates can be found in Appendix B.
+
+# 3.2 Dialogue Collection
+
+Collection Setting. Following the setting of Jang et al. (2022), annotators are instructed to make a dialogue by considering persona and corresponding knowledge under the single-person setup. In this way, one person can better understand what persona to ask as the human and what knowledge to use as the system, in comparison with two independent persons playing two roles separately. During the collection, each annotator first selects a suitable persona and then optionally identifies relevant knowledge, giving a knowledge-enhanced and persona-consistent response at last.
+
+Training and Pilot Annotation. All annotators are first required to take a training tutorial to learn the annotation procedure, requirements, and examples of annotated dialogues. Afterward, they are given 30 personas to make 10 dialogues. We provide corresponding feedback to help them adjust their annotation criteria. To establish the necessity of persona-knowledge dependency, we consider the situation where the response will be personainconsistent without the assistance of knowledge. To this end, annotators are requested to ask questions centered on the implications based on knowledge and persona. For example, the annotator is supposed to ask "Which province are you from?" instead of "Which province does Shenzhen belong to?", given the persona "I live in Shenzhen" and corresponding knowledge "Shenzhen belongs to Guangdong province".
+
+Batch Collection. After pilot annotation, we conduct dialogue collection batch by batch and regularly coach the quality of collected data<sup>7</sup>. For each batch, we sample personas different from previously annotated dialogues to increase its diversity in the whole dataset. The addition, deletion, and revision of persona and knowledge are also accepted and updated at the batch level<sup>8</sup>.
+
+Grounding Annotation. We also gather the labels of grounding knowledge sources for the system's responses by asking the annotators to specify the sources they draw from while providing responses, such as PERSONA or DOCUMENTS. For instance, gen
+
+<table><tr><td>KBP</td><td>Train</td><td>Valid</td><td>Test</td></tr><tr><td># dialogues</td><td>1,981</td><td>248</td><td>248</td></tr><tr><td># samples</td><td>9,821</td><td>1,227</td><td>1,229</td></tr><tr><td># avg turns</td><td>4.96</td><td>4.93</td><td>4.96</td></tr><tr><td># utterances</td><td>19,642</td><td>2,454</td><td>2,458</td></tr><tr><td># avg length</td><td>17.6</td><td>17.3</td><td>17.5</td></tr><tr><td># resp w/ persona</td><td>86.1%</td><td>84.4%</td><td>85.3%</td></tr><tr><td># resp w/ p_and_k</td><td>76.3%</td><td>74.2%</td><td>75.1%</td></tr></table>
+
+Table 1: Statistics of KBP dataset.
+
+erating a response may rely on persona alone or both persona and knowledge. With the labels of these grounded sources, the planning abilities of the dialogue systems can be quantitatively measured.
+
+# 3.3 Statistical Analysis
+
+We organize the collected personas (PERSONA source), persona-related knowledge $^{9}$  (DOCUMENTS source), and dialogues in the form shown in Figure 1(c). We finally collect 2,477 dialogues and 24,554 utterances with 5 turns per dialogue on average. We then split the collected data into train, validation, and test sets using the 8:1:1 ratio. The dataset statistics are summarized in Table 1, including the number of dialogues, utterances, average length, as well as data sources used. The average length per utterance reaches 17.6, hinting at the informativity and depth of the conversation. It is shown that over  $86\%$  of responses used persona (i.e., resp w/persona) and  $76\%$  used both persona and knowledge (i.e., resp w/p_and_k), which shows that KBP is capable as a benchmark to evaluate the different grounding abilities of models.
+
+# 4 Method
+
+# 4.1 Task Definition
+
+We first provide a general definition of a dialogue system that requires multiple sources and then we instantiate the definition in the context of personalized knowledge-grounded dialogue. For each dialogue, the dialogue context  $\mathbf{c} = \{u_1, s_1, u_2, s_2, \dots, u_t\}$  and different knowledge sources  $\mathbf{K} = \{K_1, K_2, \dots, K_i\}$  are provided, where  $K_i = \{k_i^1, k_i^2, \dots, k_i^j\}$  indicates the  $i_{th}$  source's name of  $\mathbf{K}$ .  $k_i^j$  denotes the  $j_{th}$  knowledge in natural language from  $K_i$ . If  $K_2$  is reliant on  $K_1$ , knowledge should be retrieved from  $K_2$  based on the selected knowledge in  $K_1$ . Such reliance should also be embodied in the construction
+
+of  $K_{2}$ , in a way such as  $K_{2} = \{k_{1}^{1}:\{k_{2}^{1},k_{2}^{2}\} ,k_{1}^{2}:$ $\{k_2^3,k_2^4\} ,\ldots \}$ . The goal of the system is to generate a response  $s_t$  conditioned on  $c$  and a set of knowledge  $\{k_i^j,\dots ,k_n^m\}$  retrieved from  $\pmb{K}$  if required<sup>10</sup>. Specifically in the context of personalized knowledge-grounded dialogue, we regard  $\{\mathrm{PERSONA, DOCUMENTS}\}$  as  $\{K_{1},K_{2}\}$  respectively. There is a potential dependency between these two sources, and the goal is to generate a response  $s_t$  conditioned on a set of knowledge  $\{p_i^j,\dots k_m^n\}$ , where  $p_i^j$  is retrieved from PERSONA and  $k_{m}^{n}$  is retrieved from DOCUMENTS. The response can also be generated conditioned on a set of knowledge from a single source PERSONA or without any sources.
+
+# 4.2 Supervised SAFARI
+
+There are three different steps in our proposed SA-FARI framework: Planning, Retrieval, and Assembling as shown in Figure 1(b). We will introduce them step-by-step under both supervised and unsupervised settings.
+
+Planning The goal of the planning step is to make a series of decisions to decide whether or not the corresponding source of knowledge is required and determine their call order if needed. Since the dependency relationship is previously known, we only need to make sure that a certain knowledge source is called after the sources it depends on. Thus, we formulate this task as sequence-to-sequence generation by directly outputting either required sources in execution order or NULL if the response does not need any knowledge as follows:
+
+$$
+\mathcal {M}: \boldsymbol {c} \rightarrow K _ {i}, K _ {j}, \dots , K _ {n} \quad o r \quad \text {N U L L}, \tag {1}
+$$
+
+where  $\mathcal{M}$  is parameterized by LLMs. We add  $K_{i},\ldots ,K_{n}$  and NULL into the vocabulary of LLMs as special tokens. The key idea here is similar to the recent ToolkenGPT (Hao et al., 2023), which regards different tools as special tokens in the vocabulary. Besides that, we add other special tokens to indicate the different parts of the input, i.e., [SOURCE] and [EOS] to indicate the start and end positions of sources. In this way, LLM can model the dependency between different sources and learn when and how to call certain sources.
+
+Retrieval According to the output of the last step, there are two cases in this step: (1) the response does not need any external sources of knowledge,
+
+![](images/3ce83cf4a6da8f126c52f4fb054551635fc1e775c2409142653c6ac35902893d.jpg)  
+Figure 2: The supervised framework of SAFARI for personalized knowledge-grounded dialogues. We use different colors to indicate different steps. The black arrow denotes the flow of data without the involvement of LLM.
+
+and the agent can skip this step; (2) the response needs multiple sources of knowledge, and the agent strictly follows the output source order to retrieve top- $n$  related knowledge  $k_{i}^{*}$  for the  $i_{th}$  knowledge source according to the dialogue context  $\pmb{c}$ , and if there is a dependency here, it will use preceding retrieved results  $k_{j}^{*}$  in the planned execution order as a filter. Specifically, assuming the output order is PERSONA, DOCUMENTS in the planning step for a persona-consistent dialogue system, we first retrieve top-1 result  $p^{*}$  from PERSONA, and then we retrieve  $k^{*}$  from DOCUMENTS according to  $\pmb{c}$  and  $p^{*}$ . Here the utilization of  $p^{*}$  depends on the systematic design. For example, if there is a designated source  $K^{*}$  for each  $p^{*}$  (a.k.a dependency), we can simply retrieve  $k^{*}$  from  $K^{*}$ . And there is another case where all knowledge is stored together and we can concatenate  $\pmb{c}$  and  $p^{*}$  as a query in the retriever.
+
+$$
+\mathcal {R}: K _ {i}, K _ {j}, \dots , K _ {n} \rightarrow k _ {i} ^ {j}, \dots , k _ {n} ^ {m} \tag {2}
+$$
+
+Assembling We concatenate all preceding results together with the dialogue context  $c$  to generate the response:
+
+$$
+\mathcal {M}: I n p \rightarrow s _ {t}, \tag {3}
+$$
+
+where  $\text{Inp} = \{c$  [SOURCE]  $K_{i}, \dots, K_{n}$  [EOS] [MIDDLE]  $k_{i}^{j}, \dots, k_{n}^{m}$  [EOM]}. [MIDDLE] and [EOM] represent the start and end positions of retrieved results. Forming the input in this way has two advantages. Firstly, the name of the sources indicates the type of results retrieved, which provides more signals to the LLMs. Secondly, it allows us to train the language model in a multi-task manner using teacher forcing. The loss is only calculated on tokens related to the planning and the response as shown in Figure 2. We first predict the planning
+
+and then generate the response according to the preceding results when inference.
+
+# 4.3 Unsupervised SAFARI
+
+Inspired by the recent progress using LLMs as a controller to plan a call order of different models (Shen et al., 2023), we adopt a similar way here by providing detailed prompts to leverage the LLMs' capability to address the dependency between different knowledge sources. We consider two settings: zero-shot and in-context learning for planning and assembling steps here since the retrieval step is the same as above.
+
+There are different knowledge bases storing relevant information:
+
+K_1: {K_1_DESC}  
+K_2: {K_2_DESC}
+
+There exists a dependency between these knowledge bases. {DEPENDENCY_DESC}
+
+Here is the dialogue between the user and the system: {DI-ALOGUE}
+
+Based on the user's last question, please determine if it requires invoking the corresponding knowledge base. If the invocation is necessary, output the names of the knowledge bases in the order they should be invoked. If no invocation is needed, output NULL.
+
+Table 2: The zero-shot prompt of unsupervised SAFARI at planning step (translated from Chinese to English).
+
+The dialogue is as follows:
+
+{DIALOGUE}
+
+The following knowledge is retrieved from different sources of knowledge bases:  
+{MIDDLE_results}
+
+Please play the role of the system and generate a reply according to the context of the dialogue and given knowledge. Please make sure your reply is consistent with the given knowledge. If the provided knowledge is NULL, generate a response solely based on the dialogue context.
+
+System:
+
+Table 3: The zero-shot prompt of unsupervised SAFARI at assembling step (translated from Chinese to English).
+
+Planning. Instead of directly providing supervision signals, we provide a description for each source of knowledge, accompanied by the corresponding dependency between the sources. The prompts are shown in Table 2.
+
+Assembling. We feed the dialogue content and the retrieved knowledge into the prompt as organized in Table 3, adapting LLMs to generate responses according to dialogue context and retrieved results.
+
+The full prompts of the unsupervised planning step can be found in Appendix D. For few-shot in-context learning, we prepend three corresponding
+
+demonstrations from the train set to the zero-shot prompts during evaluation.
+
+# 5 Experiments
+
+# 5.1 Experimental Setups
+
+Implementation Details. We mainly choose BELLE-LLAMA-7B-2M (Ji et al., 2023) and ChatGLM-6B (Du et al., 2022; Zeng et al., 2023) as two backbone models for supervised setting since they are two popular open-source Chinese models. And we additionally add ChatGPT (gpt-3.5-turbo-0301)<sup>11</sup> for the unsupervised setting. For training, we set the batch size as 8, train models with 3 epochs and save the checkpoint with the lowest validation loss. For other hyper-parameter settings, we mainly follow the corresponding official code<sup>12</sup>. Due to the computation limit, we conduct training with LoRA (Hu et al., 2021) at one single 3090 GPU, and it cost about 4-6 hours. For the unsupervised setting, we set both the temperature and top p as 0.1 to reduce the randomness of LLMs. Besides that, we use three types of retrievers including both sparse and dense retrieval: BM25 (Robertson and Zaragoza, 2009), DPR (Karpukhin et al., 2020), and RocketQAv2 (Ren et al., 2021). We only retrieve the top-ranked result from each source in the experiments<sup>13</sup>.
+
+Evaluation Metrics. During the evaluation, we use different metrics at different steps. We use F1 for planning, Recall@1 for retrieval, and BLEU (Papineni et al., 2002), Rouge-L (Lin, 2004). For assembling, Knowledge Consistency (K.C) and Persona Consistency (P.C) are calculated using our finetuned NLI models (Madotto et al., 2019). More details including retrieval models, NLI models, and NLI metrics can be found in Appendix C.
+
+# 5.2 Performance of Planning
+
+There are three types of decisions representing different sources required in the next step: NULL, PERSONA, and Both (selecting both PERSONA and DOCUMENTS). Table 4 demonstrates the F1 of planning under different settings. Under supervised settings, despite LLMs achieving high F1 scores at Both, the performance at NULL and Persona is still unsatisfactory, since there are fewer training samples in these two cases. On the other hand, under
+
+11https://openai.com/blog/chatgpt  
+$^{12}$ https://github.com/THUDM/ChatGLM-6B and https://github.com/LianjiaTech/BELLE  
+13The effects of different choices is analyzed at Section 6.
+
+<table><tr><td>Model</td><td>NULL</td><td>Persona</td><td>Both</td></tr><tr><td colspan="4">Supervised</td></tr><tr><td>BELLE-LLAMA-7B-2M</td><td>42.67 (194)</td><td>14.08 (17)</td><td>83.77 (1018)</td></tr><tr><td>CHATGLM-6B</td><td>47.10 (129)</td><td>31.96 (69)</td><td>86.59 (1031)</td></tr><tr><td colspan="4">Unsupervised</td></tr><tr><td>Zero-shot</td><td></td><td></td><td></td></tr><tr><td>BELLE-LLAMA-7B-2M</td><td>28.55 (940)</td><td>8.94 (54)</td><td>32.47 (235)</td></tr><tr><td>CHATGLM-6B</td><td>25.60 (1225)</td><td>0.0 (0)</td><td>0.43 (4)</td></tr><tr><td>CHATGPT</td><td>11.45 (116)</td><td>20.67 (233)</td><td>74.88 (880)</td></tr><tr><td>In-context</td><td></td><td></td><td></td></tr><tr><td>BELLE-LLAMA-7B-2M</td><td>9.22 (36)</td><td>18.21 (1193)</td><td>0.0 (0)</td></tr><tr><td>CHATGLM-6B</td><td>25.67 (1190)</td><td>1.49 (9)</td><td>4.62 (30)</td></tr><tr><td>CHATGPT</td><td>27.95 (699)</td><td>23.14 (238)</td><td>41.98 (292)</td></tr></table>
+
+unsupervised settings, the LLMs are over-confident in their decisions to use NULL, and they misunderstand the dependency between different sources (sometimes deciding to only use DOCUMENTS without PERSONA) $^{14}$ . This result reveals the LLMs' low accuracy in expressing uncertainty and fetching unknown knowledge. Furthermore, in-context learning cannot improve this situation, which is similar to the observation in Amayuelas et al. (2023).
+
+# 5.3 Performance of Retrieval
+
+With the ground-truth planning labels (except NULL), we examine three types of retrievers, including BM25, RocketQAv2, and DPR, to evaluate the retrieval performance. Table 5 presents the Recall@1 (R@1) of the different retrievers. We found that the DPR and RocketQAv2 can achieve over  $80\%$  R@1 when retrieving from PERSONA source while only about  $50\%$  from DOCUMENTS and the R@1 at DOCUMENTS† further decreases after removing the dependency. First, the semantics between different knowledge from DOCUMENTS with the dependency are similar to the same underlying persona  $p^*$ , making them more difficult to be distinguished. In addition, noisy knowledge sentences are introduced since there exists no dependency. Moreover, we observe that DPR performs the best out of these three retrievers in all sources of knowledge while BM25 performs worst<sup>15</sup>, revealing the importance of dense retrieval models in this task.
+
+Table 4: The F1 of different decisions in Planning of different LLMs under supervised/unsupervised settings. We also report the frequency of different decisions in the bracket. There are 181 NULL, 125 PERSONA and 923 PERSONA, and DOCUMENTS in the ground planning.  
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Persona</td><td colspan="3">Both</td></tr><tr><td>PERSONA</td><td>DOCUMENTS</td><td>DOCUMENTS†</td></tr><tr><td>BM25</td><td>36.80</td><td>48.97</td><td>15.05</td><td>11.37</td></tr><tr><td>RocketQAv2</td><td>80.00</td><td>92.31</td><td>50.49</td><td>35.75</td></tr><tr><td>DPR</td><td>83.20</td><td>93.07</td><td>51.67</td><td>39.33</td></tr></table>
+
+Table 5: The performance of Retrieval of different types of retrievers. There are 125 examples that only require PERSONA and 923 require both PERSONA and KNOWLEDGE. We also report the Recall@1 of DOCUMENTS without dependency (DOCUMENTS†).
+
+Therefore, we set DPR as the retriever in our experiments afterward.
+
+# 5.4 Performance of Assembling
+
+Table 6 demonstrates the performance of response generation under both supervised and unsupervised settings. Referring to Table 4, the performance of the planning step largely affects the results in the assembling step, when the retriever is the same. Mostly, better planning leads to better responses in all metrics. The supervised models are much better than unsupervised models since their planning results are much better, while ChatGPT performs best under unsupervised settings due to a similar reason. We found that BELLE achieves higher BLEU1 and Rouge-L, K.C but lower P.C than ChatGLM since the planning gap between them mainly comes from PERSONA. In addition, due to poor retrieval performance at DOCUMENTS (Table 5), the consistency score K.C is also much lower than P.C. With demonstrations in the prompt, we observe generally better performance on most metrics, since LLMs tend to accept the personalized role, rather than generating responses like "As an AI language model, I do not have persona ....". Overall, we conclude that the grounding ability of supervised models is much better than unsupervised ones, and ChatGPT performs best under the unsupervised setting.
+
+# 6 Analysis
+
+In this section, we analyze the effects of different components and the choice of the number of retrieved results, based on ChatGLM under the supervised setting. In addition, we conduct human evaluations to verify the quality of automatic evaluations.
+
+# 6.1 Impacts of Different Steps
+
+We investigate the effects of individual steps by providing the model ground-truth labels from each step to generate the response, enabling us to an-
+
+<table><tr><td>Model</td><td>BLEU1</td><td>Rouge-L</td><td>P.C</td><td>K.C</td></tr><tr><td colspan="5">Supervised Setting</td></tr><tr><td>BELLE-LLAMA-7B-2M</td><td>30.48</td><td>34.61</td><td>75.34</td><td>46.62</td></tr><tr><td>CHATGLM-6B</td><td>23.81</td><td>26.70</td><td>76.99</td><td>42.39</td></tr><tr><td colspan="5">Unsupervised Setting</td></tr><tr><td>Zero-shot</td><td></td><td></td><td></td><td></td></tr><tr><td>BELLE-LLAMA-7B-2M</td><td>11.84</td><td>19.24</td><td>30.59</td><td>27.34</td></tr><tr><td>CHATGLM-6B</td><td>6.18</td><td>14.50</td><td>14.73</td><td>24.73</td></tr><tr><td>CHATGPT</td><td>12.06</td><td>24.44</td><td>73.47</td><td>38.00</td></tr><tr><td>In-context</td><td></td><td></td><td></td><td></td></tr><tr><td>BELLE-LLAMA-7B-2M</td><td>19.51</td><td>22.25</td><td>72.98</td><td>24.89</td></tr><tr><td>CHATGLM-6B</td><td>13.74</td><td>19.69</td><td>16.92</td><td>24.89</td></tr><tr><td>CHATGPT</td><td>16.03</td><td>25.62</td><td>46.38</td><td>35.56</td></tr></table>
+
+alyze and understand the specific effects of each step in a clear and systematic way. Table 7 presents the results. First, we note that the inclusion of ground-truth planning labels or knowledge casts a positive impact on performance. Planning primarily enhances P.C and K.C, while grounding knowledge contributes to BLEU1 and Rouge-L scores. The best results are obtained when both signals are combined. Secondly, we also conduct an ablation study by removing some modules: 1) removing dependency information (-Dependency); 2) removing DOCUMENTS and only using PERSONA (- Documents); and 3) removing the planning step by always selecting PERSONA (-Planning*) or always selecting PERSONA and DOCUMENTS (-Planning**) for each turn. It can be found at the bottom of Table 7 that all metrics are dropped differently after removing different components except for Rouge-L when always selecting two knowledge sources. To conclude, SAFARI can effectively incorporate multiple sources (compared with -Documents) and further address dependency issues (compared with - Dependency). Moreover, SAFARI demonstrates its versatility by effectively handling multiple sources and efficiently selecting relevant ground knowledge. Notably, SAFARI outperforms existing methods that indiscriminately utilize all available sources (compared with -Planning**).
+
+# 6.2 Different Numbers of Retrieved Results
+
+The number of retrieved results plays a key role in the response generation. There is a trade-off between accuracy and recall, while a small number of retrieved results may not cover enough semantics but a large number may introduce additional noises. Table 8 presents the results of the different numbers of retrieved results. We observe that the
+
+Table 6: The performance of Assembling under supervised/unsupervised settings.  
+
+<table><tr><td>Model</td><td>BLEU1</td><td>RougeL</td><td>P.C</td><td>K.C</td></tr><tr><td>CHATGLM-6B</td><td>23.81</td><td>26.70</td><td>76.99</td><td>42.39</td></tr><tr><td>+ Ground Planning</td><td>24.29</td><td>27.01</td><td>86.16</td><td>57.12</td></tr><tr><td>+ Ground Retrieval</td><td>25.86</td><td>29.15</td><td>79.52</td><td>53.95</td></tr><tr><td>+ Ground P &amp; R</td><td>25.71</td><td>29.43</td><td>90.56</td><td>72.99</td></tr><tr><td>- Dependency</td><td>23.32</td><td>25.53</td><td>75.67</td><td>38.49</td></tr><tr><td>- Documents</td><td>23.06</td><td>25.34</td><td>75.91</td><td>36.53</td></tr><tr><td>- Planning*</td><td>23.51</td><td>25.98</td><td>72.90</td><td>24.89</td></tr><tr><td>- Planning**</td><td>23.69</td><td>26.81</td><td>71.60</td><td>34.91</td></tr></table>
+
+Table 7: Ablation study on the impact of different steps and modules in SAFARI.  
+
+<table><tr><td rowspan="2">Number</td><td colspan="4">Assembling</td></tr><tr><td>BLEU1</td><td>RougeL</td><td>P.C</td><td>K.C</td></tr><tr><td>1</td><td>23.81</td><td>26.70</td><td>76.99</td><td>42.39</td></tr><tr><td>2</td><td>22.70</td><td>25.57</td><td>71.03</td><td>29.45</td></tr><tr><td>3</td><td>20.69</td><td>24.05</td><td>69.73</td><td>27.91</td></tr></table>
+
+Table 8: The performance of Assembling of different number of retrieved results.
+
+performance of response generation decreases with the number, which indicates that noisy knowledge will harm the quality of the generated responses.
+
+# 6.3 Human Evaluation
+
+Human evaluation is conducted to evaluate the quality of generated response in terms of three metrics: coherence score (Coh.), persona consistency score (Per.Cs), and knowledge consistency score (Know.Cs). We randomly sample 100 responses with grounding information for each model and ask three annotators to indicate its coherence score (1-5) and whether the response is consistent with the given persona (1/0), and knowledge (1/0). Table 9 shows the result. We observe that supervised methods achieve higher performance than unsupervised ones, which corroborates the findings of the automatic evaluation results presented in Table 6. Besides that, we found BELLE achieves the highest performance across all metrics and outperforms ChatGLM since the effects of planning are not considered during human evaluation. Moreover, we also found that in-context learning brings a lower rejection rate and more human-like responses. Specifically, the rejection rate of BELLE under the setting of zero-shot learning is about  $32\%$ , while the number is reduced to  $12\%$  under in-context learning<sup>16</sup>.
+
+# 7 Conclusion
+
+In this paper, we propose a novel framework SA-FARI to incorporate multiple sources of knowledge
+
+<table><tr><td>Model</td><td>Coh.</td><td>Per.Cs (%)</td><td>Know.Cs (%)</td></tr><tr><td colspan="4">Supervised Setting</td></tr><tr><td>BELLE-LLAMA-7B-2M</td><td>4.38</td><td>72.0</td><td>63.8</td></tr><tr><td>CHATGLM-6B</td><td>4.06</td><td>68.0</td><td>59.1</td></tr><tr><td colspan="4">Unsupervised Setting</td></tr><tr><td>Zero-shot</td><td></td><td></td><td></td></tr><tr><td>BELLE-LLAMA-7B-2M</td><td>2.84</td><td>24.7</td><td>19.5</td></tr><tr><td>CHATGLM-6B</td><td>2.58</td><td>17.0</td><td>14.8</td></tr><tr><td>CHATGPT</td><td>4.00</td><td>63.4</td><td>33.3</td></tr><tr><td>In-context</td><td></td><td></td><td></td></tr><tr><td>BELLE-LLAMA-7B-2M</td><td>3.36</td><td>40.0</td><td>21.7</td></tr><tr><td>CHATGLM-6B</td><td>2.88</td><td>32.0</td><td>28.9</td></tr><tr><td>CHATGPT</td><td>4.03</td><td>54.0</td><td>48.8</td></tr></table>
+
+Table 9: The results of human evaluation. The inter-agreement is about  $86\%$ .
+
+bases and further address the dependency issue between them. Unlike previous works, SAFARI can be extended to multiple sources easily and it can handle cases that do not require any sources or require some instead of all sources between them. We build the first personalized knowledge-grounded dialogue (KBP) dataset, and experimental results prove the effectiveness and robustness of SAFARI.
+
+# Limitations
+
+In this paper, we propose the SAFARI framework to incorporate multiple sources and address the dependency issue between them by leveraging the exceptional capability of LLMs. However, we acknowledge two major limitations of the paper:
+
+SAFARI There is an error propagation issue when decoupling the source selection from the response generation. The cascaded design may propagate the error in the intermediate step. Once the planning sources are wrong, the retriever can not retrieve correct results from the knowledge bases.
+
+KBP Although the constructed dataset has already considered three different cases in which responses require 0, 1, and 2 knowledge sources, there are other useful sources in the knowledge-grounded dialogue system such as the users' memory source or other sources in different applications. Besides that, the world knowledge may become outdated as time goes by.
+
+# Ethics Statement
+
+In this paper, there are two ethical issues about the LLMs and dataset respectively.
+
+Usages of LLMs We strictly follow the license and policy of released LLMs, and we do not guarantee the content generated content by LLMs is
+
+safe and harmless. We note that LLMs may inherit hallucination issues as shown in the planning analysis, and it will plan not to use corresponding sources due to poor performance to express uncertainty. The calls of the OpenAI API in this paper were carried out by Dr. Yang Deng, one of the corresponding authors from the National University of Singapore.
+
+Human Annotation The human inspection and annotation were conducted by a reputable data annotation company, and the annotators are compensated fairly based on the market price without revealing any personal information. Besides that, our dataset may contain biased opinions due to the subjectivity of manual annotation.
+
+# Acknowledgement
+
+We would like to express our heartfelt gratitude to all anonymous reviewers for their insightful comments and suggestions. The work described here was partially supported by grants from the RGC General Research Funding Scheme (GRF) 14222922 (CUHK 2151185).
+
+# References
+
+Alfonso Amayuelas, Liangming Pan, Wenhu Chen, and William Wang. 2023. Knowledge of knowledge: Exploring known-unknowns uncertainty with large language models.  
+Yu Cao, Wei Bi, Meng Fang, Shuming Shi, and Dacheng Tao. 2022. A model-agnostic data manipulation method for persona-based dialogue generation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7984–8002, Dublin, Ireland. Association for Computational Linguistics.  
+Liang Chen, Hongru Wang, Yang Deng, Wai Chung Kwan, Zezhong Wang, and Kam-Fai Wong. 2023. Towards robust personalized dialogue generation via order-insensitive representation regularization. In Findings of the Association for Computational Linguistics: ACL 2023, pages 7337-7345, Toronto, Canada. Association for Computational Linguistics.  
+Yang Deng, Wenqiang Lei, Lizi Liao, and Tat-Seng Chua. 2023a. Prompting and evaluating large language models for proactive dialogues: Clarification, target-guided, and non-collaboration. CoRR, abs/2305.13626.  
+Yang Deng, Yaliang Li, Wenxuan Zhang, Bolin Ding, and Wai Lam. 2022. Toward personalized answer generation in e-commerce via multi-perspective preference modeling. ACM Trans. Inf. Syst., 40(4):87:1-87:28.
+
+Yang Deng, Wenxuan Zhang, Weiwen Xu, Wenqiang Lei, Tat-Seng Chua, and Wai Lam. 2023b. A unified multi-task learning framework for multi-goal conversational recommender systems. ACM Trans. Inf. Syst., 41(3):77:1-77:25.  
+Yang Deng, Wenxuan Zhang, Yifei Yuan, and Wai Lam. 2023c. Knowledge-enhanced mixed-initiative dialogue system for emotional support conversations. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, pages 4079-4095.  
+Emily Dinan, Stephen Roller, Kurt Shuster, Angela Fan, Michael Auli, and Jason Weston. 2019. Wizard of wikipedia: Knowledge-powered conversational agents.  
+Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. 2022. Glm: General language model pretraining with autoregressive blank infilling. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 320-335.  
+Tingchen Fu, Xueliang Zhao, Chongyang Tao, Ji-Rong Wen, and Rui Yan. 2022. There are a thousand hamlets in a thousand people's eyes: Enhancing knowledge-grounded dialogue with personal memory. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3901-3913, Dublin, Ireland. Association for Computational Linguistics.  
+Yunfan Gao, Tao Sheng, Youlin Xiang, Yun Xiong, Haofen Wang, and Jiawei Zhang. 2023. Chat-rec: Towards interactive and explainable llms-augmented recommender system.  
+Shibo Hao, Tianyang Liu, Zhen Wang, and Zhiting Hu. 2023. Toolkengpt: Augmenting frozen language models with massive tools via tool embeddings.  
+Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models.  
+Minlie Huang, Xiaoyan Zhu, and Jianfeng Gao. 2020. Challenges in building intelligent open-domain dialog systems. ACM Trans. Inf. Syst., 38(3):21:1-21:32.  
+Yoonna Jang, Jungwoo Lim, Yuna Hur, Dongsuk Oh, Suhyune Son, Yeonsoo Lee, Donghoon Shin, Seungryong Kim, and Heuiseok Lim. 2022. Call for customized conversation: Customized conversation grounding persona and knowledge. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 10803-10812.  
+Yunjie Ji, Yong Deng, Yan Gong, Yiping Peng, Qiang Niu, Lei Zhang, Baochang Ma, and Xiangang Li. 2023. Exploring the impact of instruction data scaling on large language models: An empirical study on real-world use cases. CoRR, abs/2303.14742.
+
+Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering.  
+Hyunwoo Kim, Byeongchang Kim, and Gunhee Kim. 2020. Will I sound like me? improving persona consistency in dialogues through pragmatic self-consciousness. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 904-916, Online. Association for Computational Linguistics.  
+Minghao Li, Feifan Song, Bowen Yu, Haiyang Yu, Zhoujun Li, Fei Huang, and Yongbin Li. 2023. Api-bank: A benchmark for tool-augmented llms.  
+Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain. Association for Computational Linguistics.  
+Qian Liu, Yihong Chen, Bei Chen, Jian-Guang Lou, Zixuan Chen, Bin Zhou, and Dongmei Zhang. 2020. You impress me: Dialogue generation via mutual persona perception. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1417-1427, Online. Association for Computational Linguistics.  
+Yifan Liu, Wei Wei, Jiayi Liu, Xianling Mao, Rui Fang, and Dangyang Chen. 2022a. Improving personality consistency in conversation by persona extending. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, pages 1350-1359.  
+Yifan Liu, Wei Wei, Jiayi Liu, Xianling Mao, Rui Fang, and Dangyang Chen. 2022b. Improving personality consistency in conversation by persona extending. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management. ACM.  
+Andrea Madotto, Zhaojiang Lin, Chien-Sheng Wu, and Pascale Fung. 2019. Personalizing dialogue agents via meta-learning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5454-5459, Florence, Italy. Association for Computational Linguistics.  
+Bodhisattwa Prasad Majumder, Harsh Jhamtani, Taylor Berg-Kirkpatrick, and Julian McAuley. 2020. Like hiking? you probably enjoy nature: Personagrounded dialog with commonsense expansions. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9194-9206, Online. Association for Computational Linguistics.  
+Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, Xu Jiang, Karl Cobbe, Tyna Eloundou, Gretchen Krueger, Kevin Button, Matthew Knight, Benjamin
+
+Chess, and John Schulman. 2022. Webgpt: Browser-assisted question-answering with human feedback.  
+Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics.  
+Joon Sung Park, Joseph C. O'Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. 2023. Generative agents: Interactive simulacra of human behavior.  
+Ruiyang Ren, Yingqi Qu, Jing Liu, Wayne Xin Zhao, QiaoQiao She, Hua Wu, Haifeng Wang, and Ji-Rong Wen. 2021. RocketQAv2: A joint training method for dense passage retrieval and passage re-ranking. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2825-2835, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Stephen Robertson and Hugo Zaragoza. 2009. The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends® in Information Retrieval, 3(4):333-389.  
+Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: Language models can teach themselves to use tools.  
+Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. 2023. Hugging-gpt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580.  
+Haoyu Song, Yan Wang, Kaiyan Zhang, Wei-Nan Zhang, and Ting Liu. 2021. BoB: BERT over BERT for training persona-based dialogue models from limited personalized data. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 167-177, Online. Association for Computational Linguistics.  
+Zhiqing Sun, Xuezhi Wang, Yi Tay, Yiming Yang, and Denny Zhou. 2023. Recitation-augmented language models.  
+Hongru Wang, Mingyu Cui, Zimo Zhou, and Kam-Fai Wong. 2022. TopicRefine: Joint topic prediction and dialogue response generation for multi-turn end-to-end dialogue system. In Proceedings of the 5th International Conference on Natural Language and Speech Processing (ICNLSP 2022), pages 19-29, Trento, Italy. Association for Computational Linguistics.  
+Hongru Wang, Huimin Wang, Lingzhi Wang, Minda Hu, Rui Wang, Boyang Xue, Hongyuan Lu, Fei Mi, and Kam-Fai Wong. 2023a. Tpe: Towards better
+
+compositional reasoning over conceptual tools with multi-persona collaboration.  
+Hongru Wang, Rui Wang, Fei Mi, Zezhong Wang, Ruifeng Xu, and Kam-Fai Wong. 2023b. Chain-of-thought prompting for responding to in-depth dialogue questions with llm.  
+Rui Wang, Jianzhu Bao, Fei Mi, Yi Chen, Hongru Wang, Yasheng Wang, Yitong Li, Lifeng Shang, Kam-Fai Wong, and Ruifeng Xu. 2023c. Retrieval-free knowledge injection through multi-document traversal for dialogue models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6608-6619, Toronto, Canada. Association for Computational Linguistics.  
+Charles Welch, Chenxi Gu, Jonathan K. Kummerfeld, Veronica Perez-Rosas, and Rada Mihalcea. 2022. Leveraging similar users for personalized language modeling with limited data. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1742-1752, Dublin, Ireland. Association for Computational Linguistics.  
+Sean Welleck, Jason Weston, Arthur Szlam, and Kyunghyun Cho. 2019. Dialogue natural language inference. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3731-3741, Florence, Italy. Association for Computational Linguistics.  
+Sixing Wu, Ying Li, Minghui Wang, Dawei Zhang, Yang Zhou, and Zhonghai Wu. 2021. More is better: Enhancing open-domain dialogue generation via multi-source heterogeneous knowledge. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2286-2300, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Sixing Wu, Ying Li, Ping Xue, Dawei Zhang, and Zhonghai Wu. 2022a. Section-aware commonsense knowledge-grounded dialogue generation with pretrained language model. In Proceedings of the 29th International Conference on Computational Linguistics, pages 521-531, Gyeongju, Republic of Korea. International Committee on Computational Linguistics.  
+Sixing Wu, Ying Li, Dawei Zhang, and Zhonghai Wu. 2022b. KSAM: Infusing multi-source knowledge into dialogue generation via knowledge source aware multi-head decoding. In Findings of the Association for Computational Linguistics: ACL 2022, pages 353-363, Dublin, Ireland. Association for Computational Linguistics.  
+Chen Xu, Piji Li, Wei Wang, Haoran Yang, Siyun Wang, and Chuangbai Xiao. 2022a. COSPLAY. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM.
+
+Jing Xu, Arthur Szlam, and Jason Weston. 2022b. Beyond goldfish memory: Long-term open-domain conversation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5180-5197, Dublin, Ireland. Association for Computational Linguistics.  
+Xinchao Xu, Zhibin Gou, Wenquan Wu, Zheng-Yu Niu, Hua Wu, Haifeng Wang, and Shihang Wang. 2022c. Long time no see! open-domain conversation with long-term persona memory. In *Findings of the Association for Computational Linguistics: ACL* 2022, pages 2639–2650, Dublin, Ireland. Association for Computational Linguistics.  
+Boyang Xue, Weichao Wang, Hongru Wang, Fei Mi, Rui Wang, Yasheng Wang, Lifeng Shang, Xin Jiang, Qun Liu, and Kam-Fai Wong. 2023. Improving factual consistency for knowledge-grounded dialogue systems via knowledge enhancement and alignment.  
+Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. React: Synergizing reasoning and acting in language models.  
+Wenhao Yu, Chenguang Zhu, Zaitang Li, Zhiting Hu, Qingyun Wang, Heng Ji, and Meng Jiang. 2022. A survey of knowledge-enhanced text generation. ACM Comput. Surv., 54(11s):227:1-227:38.  
+Aohan Zeng, Xiao Liu, Zhengxiao Du, Zihan Wang, Hanyu Lai, Ming Ding, Zhuoyi Yang, Yifan Xu, Wendi Zheng, Xiao Xia, Weng Lam Tam, Zixuan Ma, Yufei Xue, Jidong Zhai, Wenguang Chen, Zhiyuan Liu, Peng Zhang, Yuxiao Dong, and Jie Tang. 2023. GLM-130b: An open bilingual pre-trained model. In The Eleventh International Conference on Learning Representations (ICLR).  
+Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and Jason Weston. 2018. Personalizing dialogue agents: I have a dog, do you have pets too? In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2204-2213, Melbourne, Australia. Association for Computational Linguistics.  
+Chujie Zheng, Jinfeng Zhou, Yinhe Zheng, Libiao Peng, Zhen Guo, Wenquan Wu, Zheng-Yu Niu, Hua Wu, and Minlie Huang. 2022. CDConv: A benchmark for contradiction detection in Chinese conversations. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 18-29, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Hao Zhou, Chujie Zheng, Kaili Huang, Minlie Huang, and Xiaoyan Zhu. 2020. KdConv: A Chinese multi-domain dialogue dataset towards multi-turn knowledge-driven conversation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7098-7108, Online. Association for Computational Linguistics.
+
+Qi Zhu, Fei Mi, Zheng Zhang, Yasheng Wang, Yitong Li, Xin Jiang, Qun Liu, Xiaoyan Zhu, and Minlie Huang. 2023. KPT: keyword-guided pre-training for grounded dialog generation. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence, IAAI 2023, Thirteenth Symposium on Educational Advances in Artificial Intelligence, EAAI 2023, Washington, DC, USA, February 7-14, 2023, pages 14065-14073. AAAI Press.
+
+# A Dialogues Requiring Knowledge Dependency
+
+Table 10 shows another example of knowledge dependency. In this example, without considering the persona-knowledge reliance, the entity (proverb 瓦合), which is irrelevant to the band 草东没有派对, is retrieved. Although the retrieved knowledge is unrelated to the ground persona, ChatGPT still tries to combine these two disconnected pieces of information and generate an inconsistent response (Response w/o. Dependency).
+
+Ground Persona: "I really like 草东没有派对 (No Party for Cao Dong)"
+
+Retrieved Knowledge w/o. dependency: "瓦合 (The Clod) is a Chinese proverb that metaphorically represents a motley crowd."
+
+Ground Knowledge w. dependency: "瓦合 is a new album from the band 草东没有派对"
+
+User: Do you want to listen to 瓦合 (The Clod)?
+
+Response w/o. Dependency: I absolutely love草东没有派对！瓦合isaproverb that describes a motley crowd,representing a group of people without organization or coordination.So regarding瓦合,I can say that it represents a group of individuals lacking organization or coordination.Do you have any other questions I can assist you with?
+
+Response w. Dependency: You bet! I absolutely love草东没有派对！瓦合 is their newly released album, and I'm really looking forward to listening to the songs on it. Would you like to share with me which song you enjoy the most in 瓦合.
+
+··
+
+Table 10: An example dialogue with ChatGPT showing persona-knowledge dependency (translated from Chinese to English)
+
+# B Data Preparation
+
+Seed Preparation In the original persona dataset, DuLeMon, most of the persona does not contain much knowledge. For example, about  $30\%$  to  $40\%$  of persona is about the name such as "my name is XXX". After clustering, we removed topics that contain less than 20 personas. There are 5 topics left: career, book, music, movie, personality. Then we add some basic persona such as gender, age, locality, nation, sport, and others.
+
+Persona and Knowledge Matching For each persona description after segmentation, we first remove stop-words<sup>17</sup> and frequently used words.
+
+<table><tr><td>Attributes</td><td>Templates</td></tr><tr><td>nick_name, birth_place, ... rewards, members, ...</td><td>{head entity}’s {attribute} is {tail entity} The {attribute} of {head entity} contains {tail entity}</td></tr></table>
+
+Table 11: Two major templates to convert triples into natural language descriptions (translated from Chinese to English).
+
+And then we filter duplicated words to reduce unnecessary computation. For the left words, we map them to the corresponding knowledge base one by one. We discarded some useless attributes such as penmanship and pronunciation, and then translate the matched triples to natural language sentences according to pre-defined templates as shown in Table 11. If we do not find any matched triples, we just discard the persona. At last, we manually checked each persona description to make sure (1) there is no repetitive knowledge statement, and (2) each persona contains at least 5 knowledge statements<sup>18</sup>.
+
+# C Implementation Details
+
+We first illustrate the details of finetuning and then introduce our definition of NLI metrics: P.C and K.C.
+
+Retrieval Models. We finetune RocketQAv2 and DPR using our own KBP dataset by regarding (context, used_persona / document) as the positive and (context, unrelated_persona / document) as the negative. We set epochs as 5 and max sequence length as 512, and mainly follow the scripts: https://github.com/PaddlePaddle/RocketQA and https://github.com/Alibaba-NLP/Multi-CPR/tree/main/retrieval respectively for other parameters. For RocketQAv2, we load the weights of pre-trained model zh_durreader_de_v2 as introduced in the official homepage, which is trained on the largest Chinese QA dataset, and we use 12-layer bert-base-chinese with 110M parameters as backbone model for DPR.
+
+NLI Models. Following previous work (Kim et al., 2020; Cao et al., 2022), we finetune an NLI model (Welleck et al., 2019) using our own dataset by regarding (ground_persona / document, response) as the positive and randomly sampled (unrelated_persona / document, response) as the negative. We also use bert-base-chinese as the backbone model. We concatenate and encode the ground persona/document  $k$  and response  $r$  in the form of [CLS]  $k$  [SEP]  $r$  [SEP], and we train the
+
+$$
+\mathbf {C} (r, g) = \left\{ \begin{array}{l l} \mathbf {N L I} (r, g), & \text {i f t h e r e i s g r o u n d i n g g f o r r , a n d p l a n n i n g a l s o u s e s g ^ {\prime}}. \\ 0, & \text {i f t h e r e i s g r o u n d i n g g f o r r , a n d p l a n n i n g d o e s n o t u s e g ^ {\prime}}. \\ 0, & \text {i f t h e r e i s n o g r o u n d i n g g f o r r , a n d p l a n n i n g u s e s g ^ {\prime}}. \\ 1, & \text {i f t h e r e i s n o g r o u n d i n g g f o r r , a n d p l a n n i n g d o e s n o t u s e g ^ {\prime}}. \end{array} \right. \tag {4}
+$$
+
+model to predict whether responses are consistent with corresponding personas or documents. The batch size for fine-tuning is 8. The maximal training epoch is 5, and the max sequence length of the encoder is 512. In the experiments, we use the AdamW optimizer with a learning rate of 2e-5 and an epsilon of 1e-6. We evaluate the NLI model on the KBP test set every 500 iterations during training, and we save the checkpoint with the highest performance on the KBP test set. The fine-tuned model achieves  $>95\%$  accuracy for both persona  $(95.94\%)$  and knowledge  $(96.95\%)$  on the KBP test dataset. We then calculate the persona consistency (P.C) and knowledge consistency (K.C) according to the output of the NLI model. Compared with many previous works (Madotto et al., 2019; Cao et al., 2022) that calculate the NLI score only when ground-truth knowledge includes personas or documents, our framework is more sophisticated and introduces the planning step, which considers the situation where responses do not require any ground knowledge. Thus, if we only calculate the consistency score in the occasion where there exists ground truth personas/documents, it is unfair and inaccurate for our framework, since wrong ground knowledge could also hurt the quality of responses and the system is not penalized when it gives wrong planning (e.g. always calling external sources). We design a new calibrated metric for dialogue consistency as described in Eq 4,5,6,7:
+
+$$
+\mathbf {N L I} (r, g) = \left\{ \begin{array}{l l} 1, & \text {i f r e n t a i l s g} \\ 0, & \text {i f r d o e s n o t e n t a i l g} \end{array} \right. \tag {5}
+$$
+
+$$
+\mathbf {P . C} = \frac {\sum_ {i} ^ {m} C \left(r _ {i} , p _ {i}\right)}{m} \tag {6}
+$$
+
+$$
+\mathbf {K}. \mathbf {C} = \frac {\sum_ {i} ^ {m} C \left(r _ {i} , k _ {i}\right)}{m} \tag {7}
+$$
+
+There are four cases during experiments as shown in Eq 4: 1). there exists ground-truth grounding  $g$  for the response  $r$  in the test set and the planning also decides to retrieve ground knowledge  $g'$  from corresponding sources (either PERSONA or DOCUMENTS); 2). there is  $g$  for  $r$ , but the planning
+
+decides not to use  $g'$ ; 3). there is no grounding  $g$  for  $r$ , while the planning decides to use  $g'$ ; 4). there exists no  $g$ , and the planning step decides not to use  $g'$  either. We only calculate the NLI score of  $(g, r)$  using finetuned NLI model in the first case of Eq 4. With this definition, we can get the calibrated score P.C and K.C for the KBP test set.
+
+# D Prompts Templates under Unsupervised Setting
+
+Table 12 demonstrates the zero-shot prompt of the SAFARI planning step on the KBP dataset, and the zero-shot prompt for the assembling step is the same as Table 3.
+
+There are two knowledge bases storing relevant information:
+
+PERSONA: This knowledge base stores information related to system personas, such as gender, age, place of origin, hobbies, personality traits, and other relevant data.
+
+DOCUMENTS: This knowledge base stores domain-specific knowledge related to system personas, such as the domain knowledge about the place of origin.
+
+There exists a dependency between these knowledge bases. The invocation of DOCUMENTS relies on the results from PERSONA. Please ensure the correct order of invoking them.
+
+Here is the dialogue between the user and the system: {dialogue_history}
+
+Based on the user's last question, please determine if it requires invoking the corresponding knowledge base. If the invocation is necessary, output the names of the knowledge bases in the order they should be invoked. If no invocation is needed, output NULL.
+
+Table 12: The prompts of unsupervised SAFARI on KBP dataset (translated from Chinese to English)
+
+# E Human Evaluation
+
+We find that human is more likely to find persona-inconsistent cases. There are some responses that have intra-sentence contradictions (Zheng et al., 2022), for example, "My zodiac signs are Aries and Taurus". In addition, there are other responses related to the persona description that are inconsistent. Both these types of responses are easy to identify by humans but hard for the NLI model to detect, resulting in lower Per.Cs during human evaluation.

2023/Large Language Models as Source Planner for Personalized Knowledge-grounded Dialogues/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5d83f81cacc414e675a98796ba4d9cfc87b01f67f95d9a0a45af24632af88f47
+size 407898

2023/Large-Scale and Multi-Perspective Opinion Summarization with Diverse Review Subsets/4b542fcd-babd-4d16-9e9d-894d20925ab6_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1ae74fa1e5ff942487b388a6e6072aebd8ee18929cedd2b4fcb34481ddc27002
+size 651973

2023/Large-Scale and Multi-Perspective Opinion Summarization with Diverse Review Subsets/full.md

@@ -0,0 +1,407 @@
+# Large-Scale and Multi-Perspective Opinion Summarization with Diverse Review Subsets
+
+Han Jiang, Rui Wang, Zhihua Wei\*, Yu Li, Xinpeng Wang
+
+Department of Computer Science and Technology, Tongji University, Shanghai, China {2230780, rwang, zhihua_wei, 2331891, wangxinpeng}@tongji.edu.cn
+
+# Abstract
+
+Opinion summarization is expected to digest larger review sets and provide summaries from different perspectives. However, most existing solutions are deficient in epitomizing extensive reviews and offering opinion summaries from various angles due to the lack of designs for information selection. To this end, we propose SUBSUMM, a supervised summarization framework for large-scale multi-perspective opinion summarization. SUBSUMM consists of a review sampling strategy set and a two-stage training scheme. The sampling strategies take sentiment orientation and contrastive information value into consideration, with which the review subsets from different perspectives and quality levels can be selected. Subsequently, the summarizer is encouraged to learn from the sub-optimal and optimal subsets successively in order to capitalize on the massive input. Experimental results on AmaSum and Rotten Tomatoes datasets demonstrate that SUBSUMM is adept at generating pros, cons, and verdict summaries from hundreds of input reviews. Furthermore, our in-depth analysis verifies that the advanced selection of review subsets and the two-stage training scheme are vital to boosting the summarization performance.
+
+# 1 Introduction
+
+A plethora of online resources has been appealing for techniques of automatic information mining. Opinion summarization, as a task of generalizing views from a group of documents (e.g., reviews, posts, and discussions) related to an entity and presenting them in text form, has received considerable attention. The summarization of user opinions is of great advantage to public opinion research, marketing analysis, and decision-making (Im et al., 2021). While circumventing the tedious document-by-document browsing, it also offers more significant details compared to a single sentiment rating (Wang and Wan, 2021).
+
+Due to the burgeoning amount of online reviews and user needs, opinion summarization is expected to (1) process larger sets of documents and (2) provide summaries from different perspectives. One mainstream solution models the cross-document relations with sentence or document representations, using mean function (Chu and Liu, 2019; Bražinskas et al., 2020b; Li et al., 2020), convex combination (Iso et al., 2021), and graph (Erkan and Radev, 2004; Ganesan et al., 2010) as well as other hierarchical structures (Isonuma et al., 2021; Amplayo et al., 2021a). These approaches are proven to achieve remarkable results with a moderate amount of reviews, usually within 10 (Shapira and Levy, 2020); however, they perform unsatisfactorily when the number of reviews further increases, as they focus on the fusion rather than the selection of the information. Another solution concatenates the reviews for long-range language models (Beltagy et al., 2020; Zaheer et al., 2020; Mao et al., 2022; Pang et al., 2023) and Large Language Models (LLMs; OpenAI, 2023), which converts multi-document summarization into single-document summarization (Bražinskas et al., 2020a; Oved and Levy, 2021; Ke et al., 2022; Brazinskas et al., 2022; Bhaskar et al., 2023). Despite the benefit brought by the LLMs, these methods struggle to handle the overlong combined reviews, missing a step to select from them either. Bražinskas et al. (2021) first proposes to select smaller subsets of input reviews and provides verdict, pros, and cons summaries, yet differentiated treatments of different perspectives are not reflected in their method. Limited by data, there are seldom works targeting large-scale and multi-perspective opinion summarization.
+
+To address the problems, we propose SUBSUMM, a supervised summarization framework for large-scale and multi-perspective opinion summarization. SUBSUMM comprises a review sampling strategy set and a two-stage training scheme. The
+
+review sampling strategies are formulated with sentiment analysis and contrastive information valuation. With different strategies, the review subsets from different angles and quality levels can be selected. Then, the two-stage training method enables the summarization model to learn from the sub-optimal and optimal review subsets successively to fully utilize the input reviews within the model capacity. During the training stage II, a contrastive loss term is incorporated to further boost the performance of the summarizer.
+
+By coupling with SUBSUMM, the Pre-trained Language Model (PLM) outperforms previous state-of-the-art models and LLMs under zero-shot settings on the AmaSum and Rotten Tomatoes datasets in our experiments, which demonstrates the superiority of the proposal. Further analysis also proves the effectiveness of the two modules in SUBSUMM.
+
+The contributions of this paper are as follows.
+
+- We propose a large-scale opinion summarization framework<sup>1</sup> to address the challenge of summarizing large review sets and providing opinions from different perspectives by selecting valuable review subsets.  
+- We present (1) a review sampling strategy set based on sentiment analysis and contrastive information valuation and (2) a two-stage training scheme promoting the digestion and absorption of the massive input.  
+- We substantiate the effectiveness of the proposed opinion summarization framework SUBSUMM with sufficient experiments and in-depth analysis on two opinion summarization datasets from different domains.
+
+# 2 Related Work
+
+Opinion Summarization As high-quality annotation for the large opinion corpora is expensive to obtain (Ge et al., 2023), most works of opinion summarization are unsupervised, summarizing a limited number of reviews. Among the abstractive approaches, VAE-based and synthetic-dataset-based models have the upper hand.
+
+The VAE-based models (Chu and Liu, 2019; Bražinskas et al., 2020b; Li et al., 2020; Iso et al.,
+
+2021; Isonuma et al., 2021) summarize through the aggregation of the latent representations of the reviews. COOP (Iso et al., 2021) considers the convex combination of input review representations. These methods work well with fewer reviews, while they suffer a performance drop when processing numerous reviews.
+
+The synthetic-dataset-based methods (Amplayo and Lapata, 2020; Bražinskas et al., 2020a; Oved and Levy, 2021; Wang and Wan, 2021; Amplayo et al., 2021b; Ke et al., 2022; Brazinskas et al., 2022) transform the unsupervised task into a supervised task by constructing review-summary pairs from original data. PASS (Oved and Levy, 2021) applies systematic perturbations to the input reviews for more candidate summaries and trains a classifier to rank the candidates. CONSISTSUM (Ke et al., 2022) measures the distances between reviews from aspect, sentiment, and semantics to create highly relevant review-summary pairs. ADA-SUM (Brazinskas et al., 2022) first fine-tunes the PLM with a synthetic dataset, then performs fine-tuning in a few-shot manner. The idea of making full use of the original text is embodied thoroughly in these methods.
+
+Benefiting from the growth of annotated data for opinion summarization, there are some emergent studies on supervised methods. Bražinskas et al. (2021) provide a large-scale opinion summarization dataset enabling supervised training. They formulate the task as jointly learning to select informative reviews and summarize the opinions, and their solution SELSUM is based on reinforcement learning (REINFORCE; Williams, 1992). Aiming at avoiding the challenges brought by reinforcement learning, we decouple the process of selection and summarization in this work.
+
+Contrastive Learning Contrastive learning in automatic summarization (Cao and Wang, 2021; Xu et al., 2021; Sun and Li, 2021; Liu and Liu, 2021; Liu et al., 2022) also gives us much inspiration. CLIFF (Cao and Wang, 2021) creates negative samples with automatically generated erroneous summaries. SIMCLS (Liu and Liu, 2021) trains an extra model with contrastive learning to evaluate and rank the candidate summaries. BRIO (Liu et al., 2022) introduces contrastive learning to assign a dual role to the model, alleviating inference performance degradation. In this work, we explore contrastive learning for multi-document summarization rather than single-document sum
+
+![](images/2459afd82185adf8f667d034b6303a569fbdeab3b851aaf38d1cc84f8e0d1c6a.jpg)  
+Review subset produced by Sentiment-Random
+
+![](images/edafce765bc4dbe47669830abb9016642c7ddff7cb75c41371be3ee4f5e6b7b4.jpg)  
+Review subset produced by Sentiment-Information
+
+![](images/d2c6dd95ffedb7831ea25bc2aa20e39c7b9e183316857dd113ce09a5ca694e70.jpg)  
+(a) Review sampling strategy set
+
+![](images/4909c7780af20d523278a9d6f8d57c5fe420ee397962db1e050704235c777e31.jpg)  
+(b) Two-stage progressive training scheme  
+Figure 1: The framework of SUBSUMM. (a) The review sampling strategy set based on sentiment analysis and information valuation. The three strategies, Random Sampling, Sentiment-Random Sampling, and Sentiment-Information Ranking are abbreviated as Random, Sentiment-Random, and Sentiment-Infomation. (b) The two-stage training scheme, under which the PLM (BART (Lewis et al., 2020) in this work) learns to summarize sub-optimal and optimal subsets successively.
+
+marization, and a PLM is fine-tuned contrastively for the information valuation.
+
+# 3 Methodology
+
+We introduce SUBSUMM, a supervised framework for large-scale and multi-perspective opinion summarization, as illustrated in Fig. 1. SUBSUMM is composed of a review sampling strategy set regarding sentiment orientation and information value, as elucidated in Sec. 3.1; and a two-stage training scheme, where contrastive learning with candidate summaries is extra performed, see Sec. 3.2.
+
+Given the entity set and the corresponding sample set, for every sample  $\{R_{1:N}, S\}$ , the goal of opinion summarization is to learn a function  $f$  that takes the review set as input and outputs a summary as close as possible to the reference:
+
+$$
+S \leftarrow f \left(R _ {1: N}\right) \tag {1}
+$$
+
+where  $R_{1:N}$  is the original review set, and  $S$  is the reference opinion summary. This paper mainly discusses the situation where  $R_{1:N}$  is too large to be processed with most language models. For review set  $R_{1:N}$ , let  $R_{1:K}$  be the review subset where  $K \ll N$ .
+
+# 3.1 Review Sampling Strategy Set
+
+Sentiment Analysis We leverage sentiment analysis to filter the reviews roughly. It is supposed that the reviews with similar sentiment orientations are less likely to conflict in content than those with
+
+contrary sentiment orientations. On the other hand, the sentiment tendency of the summary can be controlled by adjusting the proportion of input reviews with different sentiments; in this way, multiple angles are formed.
+
+We formulate sentiment analysis as a text classification task. The sentiment tags of reviews in  $R_{1:N}$  are computed as:
+
+$$
+p _ {i} ^ {\text {s e n}} = \operatorname {S L M} \left(R _ {i}\right) \tag {2}
+$$
+
+$$
+S e n _ {i} = \underset {j} {\arg \max } p _ {i} ^ {s e n} (j) \tag {3}
+$$
+
+where  $SLM(\cdot)$  is a PLM with a linear classification head, and  $p_i^{sen}$  refers to the probability distribution over the sentiment classes of the review  $R_{i}$ . The class with the highest probability is taken as the sentiment tag  $Sen_{i}$ . We use the rating of each review in the dataset as the sentiment label and apply a negative log likelihood loss w.r.t. the sentiment distribution  $p_i^{sen}$ . After fine-tuning, the sentiment tags of all the reviews can be obtained.
+
+Contrastive Information Valuation Information valuation has finer granularity than sentiment analysis. Intuitively, once a subset of reviews is selected for summarization, the closer the generation is to the reference, the more valuable the information in the subset may be.
+
+Given the reference summary, the information value of an input review is tied to its similarity with the reference; ROUGE (Lin, 2004) is an appropriate metric to estimate such similarity. Thus we fit the ROUGE score of a review
+
+$R_{i} = \{r_{i}^{(1)},\dots,r_{i}^{(|R_{i}|)}\}$  by modeling the correlation between the review and the whole review set:
+
+$$
+\begin{array}{l} h _ {i} ^ {(1)}, \dots , h _ {i} ^ {(| R _ {i} |)} = E n c \left(r _ {i} ^ {(1)}, \dots , r _ {i} ^ {(| R _ {i} |)}\right) (4) \\ h _ {i} = \frac {1}{\left| R _ {i} \right|} \sum_ {k = 1} ^ {\left| R _ {i} \right|} h _ {i} ^ {(k)} (5) \\ C o r r _ {i} = \frac {1}{N - 1} \sum_ {1 \leq j \leq N, j \neq i} h _ {i} h _ {j} (6) \\ \end{array}
+$$
+
+where  $Enc(\cdot)$  is a transformer (Vaswani et al., 2017) encoder, and  $h_i^{(k)}$  denotes the last hidden state of token  $r_i^{(k)}$ . The review representation  $h_i$  is computed by averaging the last hidden states of the tokens in  $R_i$ . Corr  $_i$  is the correlation score between  $R_i$  and the review set  $R_{1:N}$ . We refer to the leave-one-out setting in unsupervised opinion summarization, computing Corr  $_i$  by the dot product of  $h_i$  and the mean representation of the others.
+
+Mindful of the volume of the original review set, it is unfeasible to fit the distribution of the ROUGE score directly or employ list-wise loss. Therefore, we resort to a contrastive margin loss:
+
+$$
+\mathcal {L} = \sum_ {i = 1} ^ {N} \sum_ {r (j) > r (i)} \max  \left(0, C o r r _ {j} - C o r r _ {i} + \lambda_ {i j}\right) \tag {7}
+$$
+
+where  $r(i)$  accounts for the ranking of  $R_{i}$  when sorted by ROUGE  $(R_{i},S)$  in descending order, and  $\lambda_{ij} = \lambda (r(j) - r(i))$  is a margin varying with the rankings, defined following Zhong et al. (2020); Liu and Liu (2021); Liu et al. (2022). The pairwise loss allows the model to learn the ROUGE rankings of a large review set. After fine-tuning, we can get the estimated information values of all the reviews.
+
+Multi-level Review Sampling Strategies Drawing support from the sentiment analysis and the contrastive information valuation, diverse review sampling strategies can be formed to select  $R_{1:K}$  out of  $R_{1:N}$ . We find that it is not ideal to accomplish the task with a single optimal subset, which will be explained in Sec. 4.3. To tackle the problem, we introduce stochastic factors to develop sampling strategies of multiple quality levels.
+
+The sampling strategy set consists of the following three strategies:
+
+- Random Sampling: Randomly sample  $K$  reviews from the original review set  $R_{1:N}$  as the subset  $R_{1:K}$ .
+
+- Sentiment-Random Sampling: Firstly, divide all reviews into positive and negative types according to their sentiment tags  $\text{Sen}_i$ . Secondly, the number of reviews of each type in  $R_{1:K}$  is determined by the type of the reference summary:
+
+$$
+\left(K ^ {+}, K ^ {-}\right) = \left\{ \begin{array}{l l} (K, 0), & \text {p r o s} \\ (0, K), & \text {c o n s} \\ \left(\frac {K N ^ {+}}{N}, \frac {K N ^ {-}}{N}\right), & \text {v e r d i c t} \end{array} \right. \tag {8}
+$$
+
+where  $(K^{+},K^{-})$ $(N^{+},N^{-})$  stands for the numbers of positive and negative reviews in  $R_{1:K}$ $R_{1:N};K^{+} + K^{-} = K,N^{+} + N^{-} = N.$  Finally, randomly sample  $K^{+},K^{-}$  reviews from the positive and negative types respectively for  $R_{1:K}$
+
+- Sentiment-Information Ranking: Firstly, compute  $(K^{+}, K^{-})$  likewise. Secondly, sort the reviews in descending order by the estimated information value  $Corr_{i}$  in two types separately. Finally, take the top- $K^{+}$  positive reviews and the top- $K^{-}$  negative reviews for  $R_{1:K}$ .
+
+The quality of the corresponding review subsets should improve in sequence.
+
+# 3.2 Two-Stage Training for Large-Scale Opinion Summarization
+
+SUBSUMM embodies a two-stage training scheme encouraging the summarizer to learn from the suboptimal and optimal review subsets successively.
+
+In stage I, we choose the sub-optimal strategy, Sentiment-Random Sampling to re-sample the review subset  $\dot{R}_{1:K}$  at each training epoch and train the model with standard maximum likelihood estimation (MLE):
+
+$$
+\theta^ {*} = \arg \max  _ {\theta} \sum_ {i} \log p _ {\theta} (S | \dot {R} _ {1: K} ^ {(i)}) \tag {9}
+$$
+
+where  $\theta$  denotes the parameters of the abstract model, and  $p_{\theta}$  represents the probability derived by the parameters. The cross entropy loss is defined over the reference sequence of length  $l$  as:
+
+$$
+\begin{array}{l} \mathcal {L} _ {\mathrm {I}} = \mathcal {L} _ {x e n t} = \\ - \sum_ {i = 1} ^ {l} \sum_ {s \in \mathcal {V}} p ^ {*} (s | \dot {R} _ {1: K}, S _ {<   i}) \log p _ {\theta} (s | \dot {R} _ {1: K}, S _ {<   i}) \tag {10} \\ \end{array}
+$$
+
+where  $s$  can be any token in the vocabulary  $\mathcal{V}$ , and  $p^*$  refers to an one-hot distribution.  $S_{<i}$  stands for a pre-defined start token and the first  $i - 1$  tokens of the reference summary. However, Standard MLE is prone to exposure bias since it heavily relies on the ground-truth sequence (Zhang et al., 2019). Meanwhile, whichever strategy is adopted, the reviews sampled are only a part of the original review set, where the information can be further exploited.
+
+In stage II, we take a cue from the practice of assigning probability mass to candidate summaries during training (Liu et al., 2022). Theoretically, assigning probability mass to a summary means an opportunity for the summary to pass on knowledge to the model through backpropagation. Hence the range of probability mass allocation is essentially the range of model learning, and better candidate summaries ought to compete for more probability mass. We plan to reuse the original review set via the candidate summaries.
+
+To start with, we slightly modify the optimal strategy (i.e., Sentiment-Information Ranking), as some perturbations are required to obtain various candidate summaries for contrastive learning:
+
+- Sentiment-Information Ranking (modified): After computing  $(K^{+}, K^{-})$  in Eq. 8, take the estimated information value  $Corr_{i}$  of each review as weight to sample  $K^{+}, K^{-}$ reviews from the positive and negative types severally.
+
+Next, the modified optimal strategy is repeatedly conducted to get  $M$  review subsets, with which  $M$  candidate summaries  $\hat{S}_1,\hat{S}_2,\dots,\hat{S}_M$  are generated by the model from stage I. The review subset produced by the original optimal strategy, denoted by  $\ddot{R}_{1:K}$ , will be the training input. We again calculate the ROUGE scores of the reviews in  $\ddot{R}_{1:K}$  with the reference summary  $S$  to derive the rankings and apply a contrastive loss term similar to Eq. 7:
+
+$$
+\mathcal {L} _ {c t r} = \sum_ {i = 1} ^ {M} \sum_ {r (j) > r (i)} \max  \left(0, L h _ {j} - L h _ {i} + \lambda_ {i j}\right) \tag {11}
+$$
+
+where  $Lh_{i}$  is the length-normalized likelihood of the candidate summary  $\hat{S}_i$ , which is defined following Liu et al. (2022):
+
+$$
+L h (S) = \frac {\sum_ {i = 1} ^ {| S |} \log p _ {\theta} \left(s _ {i} \mid \ddot {R} _ {1 : K} , S _ {<   i}\right)}{| S | ^ {\alpha}} \tag {12}
+$$
+
+Here  $\alpha$  is a length penalty hyperparameter. This term enforces the model to assign more probability mass to better candidate summaries.
+
+Finally, to maintain the generation ability of the pre-trained model, we follow Edunov et al. (2018) to use the multi-task loss:
+
+$$
+\mathcal {L} _ {\mathrm {I I}} = \mathcal {L} _ {x e n t} + \gamma \mathcal {L} _ {c t r} \tag {13}
+$$
+
+where  $\gamma$  is the weight of the contrastive loss term. By involving the candidate summaries in training, stage II raises the utilization rate of the original review set and alleviates the problem of exposure bias; it acts as a complement to stage I considering the addition of the review subsets with higher quality and the contrastive loss term.
+
+During inference, given a review set  $R_{1:N}$ , SUBSUMM predicts the sentiment tag and information value of each review with the fine-tuned PLMs in Sec. 3.1, then selects the optimal review subset  $R_{1:K}$  according to the Sentiment-Information Ranking strategy and summarizes the subset using the summarization model from stage II.
+
+# 4 Experiments
+
+# 4.1 Experimental Settings
+
+Datasets We choose two opinion summarization datasets with large review sets as our testbed. The statistics are shown in Appendix A.
+
+AmaSum $^2$  (Bražinskas et al., 2021) is a product review dataset where each sample contains a large number of reviews and reference summaries written by professional reviewers. Unlike other datasets, AmaSum provides reference summaries from three perspectives, namely verdict, which is equivalent to general opinion summary; pros and cons, which summarize the most important positive and negative details. As shown in Table 6, with  $4.2\mathrm{k}$  tokens on average, the combined reviews in AMASUM are too long to summarize with most summarizers. We refer to the preprocessing in SELSUM but split the dataset into three partitions with different targets.
+
+Rotten Tomatoes<sup>3</sup> (RT; Wang and Ling, 2016) is a large-scale movie review dataset. For each movie, a one-sentence critic consensus is constructed by an editor to summarize the opinions in professional critics, which is treated as the reference summary. We follow Amplayo et al. (2021b) to preprocess the dataset; the data in RT and the verdict partition of AmaSum are equally treated in our experiments.
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">Pros</td><td colspan="3">Cons</td><td colspan="3">Verdict</td></tr><tr><td>R-1</td><td>R-2</td><td>R-L</td><td>R-1</td><td>R-2</td><td>R-L</td><td>R-1</td><td>R-2</td><td>R-L</td></tr><tr><td>Unsupervised MEANSUM†</td><td>10.44</td><td>0.63</td><td>9.55</td><td>5.95</td><td>0.45</td><td>5.29</td><td>13.78</td><td>0.93</td><td>11.70</td></tr><tr><td>LEXRANK†</td><td>14.12</td><td>1.50</td><td>12.81</td><td>8.28</td><td>0.82</td><td>7.24</td><td>15.12</td><td>1.84</td><td>12.60</td></tr><tr><td>COPYCAT†</td><td>15.12</td><td>1.48</td><td>13.85</td><td>6.81</td><td>0.82</td><td>5.89</td><td>17.05</td><td>1.78</td><td>14.50</td></tr><tr><td>EXTSUM†</td><td>19.06</td><td>2.47</td><td>17.49</td><td>11.63</td><td>1.19</td><td>10.44</td><td>18.74</td><td>3.01</td><td>15.74</td></tr><tr><td>Supervised SELSUM†</td><td>21.29</td><td>4.00</td><td>19.39</td><td>14.96</td><td>2.60</td><td>13.07</td><td>24.33</td><td>5.29</td><td>18.84</td></tr><tr><td>LONGFORMER</td><td>22.40</td><td>4.71</td><td>15.36</td><td>14.68</td><td>2.53</td><td>11.62</td><td>22.56</td><td>4.83</td><td>17.08</td></tr><tr><td>SELSUM*</td><td>23.17</td><td>4.77</td><td>21.13</td><td>15.12</td><td>2.83</td><td>13.07</td><td>22.87</td><td>4.85</td><td>18.05</td></tr><tr><td>BRIO</td><td>25.48</td><td>4.58</td><td>23.50</td><td>16.65</td><td>2.94</td><td>14.60</td><td>24.93</td><td>4.78</td><td>19.44</td></tr><tr><td>SUBSUMM</td><td>26.25</td><td>4.96</td><td>24.18</td><td>16.72</td><td>3.00</td><td>14.80</td><td>25.36</td><td>5.04</td><td>19.58</td></tr><tr><td>Zero-Shot QG</td><td>18.40</td><td>2.21</td><td>16.02</td><td>13.27</td><td>1.39</td><td>11.61</td><td>14.75</td><td>1.23</td><td>12.06</td></tr><tr><td>CHATGPT</td><td>18.53</td><td>2.37</td><td>14.68</td><td>13.92</td><td>1.78</td><td>11.93</td><td>22.88</td><td>3.50</td><td>19.79</td></tr></table>
+
+Table 1: Automatic evaluation results on AmaSum dataset. Best models are shown in bold and 2nd best models are underlined;  $\dagger$  means that the results are copied from Bražinskas et al. (2021). We retrained SELSUM* on pros, cons, and verdict separately for a fair comparison.  
+
+<table><tr><td>Method</td><td>R-1</td><td>R-2</td><td>R-L</td></tr><tr><td colspan="4">Unsupervised</td></tr><tr><td>W2VCENT†</td><td>13.93</td><td>2.10</td><td>10.81</td></tr><tr><td>LEXRANK†</td><td>14.88</td><td>1.94</td><td>10.50</td></tr><tr><td>OPINOSIST</td><td>14.98</td><td>3.07</td><td>12.19</td></tr><tr><td>MEANSUM†</td><td>15.79</td><td>1.94</td><td>12.26</td></tr><tr><td>SNCENT†</td><td>15.90</td><td>2.01</td><td>11.74</td></tr><tr><td>BERTCENT†</td><td>17.65</td><td>2.78</td><td>12.78</td></tr><tr><td>DENOISESUM†</td><td>21.26</td><td>4.61</td><td>16.27</td></tr><tr><td colspan="4">Weakly Supervised</td></tr><tr><td>PLANSUM†</td><td>21.77</td><td>6.18</td><td>16.98</td></tr><tr><td colspan="4">Supervised</td></tr><tr><td>BRIO</td><td>23.72</td><td>5.16</td><td>18.05</td></tr><tr><td>LONGFORMER</td><td>24.96</td><td>6.34</td><td>18.40</td></tr><tr><td>SUBSUMM</td><td>24.96</td><td>6.66</td><td>19.08</td></tr><tr><td colspan="4">Zero-Shot</td></tr><tr><td>QG</td><td>18.14</td><td>2.34</td><td>14.28</td></tr><tr><td>CHATGPT</td><td>22.73</td><td>4.21</td><td>17.52</td></tr></table>
+
+Table 2: Automatic evaluation results on RT dataset. Best models are shown in bold and 2nd best models are underlined;  $\dagger$  means that the results are copied from Amplayo et al. (2021b).
+
+Baselines Concerning the baselines, we select a series of competitive models for the two datasets.
+
+On AmaSum dataset, the baselines include (1)
+
+unsupervised extractive models LEXRANK (Erkan and Radev, 2004) and EXTSUM (Bražinskas et al., 2021); (2) unsupervised abstractive models MEAN-SUM (Chu and Liu, 2019) and COPYCAT (Bražinskas et al., 2020b); (3) supervised abstractive models SELSUM, LONGFORMER (Beltagy et al., 2020), and BRIO; and (4) zero-shot solutions related to LLMs, including GPT-3.5-turbo (CHATGPT) as well as QG (Bhaskar et al., 2023) based on QF-Summ (Ahuja et al., 2022) and GPT-3 (Brown et al., 2020).
+
+On RT dataset, the extra baselines are (1) unsupervised extractive models W2VCENT (Rossiello et al., 2017), SNCENT (Amplayo and Lapata, 2020), and BERTCENT (Amplayo et al., 2021b); (2) unsupervised abstractive models OPINOSIS (Ganesan et al., 2010) and DENOISESUM (Amplayo and Lapata, 2020); (3) weakly supervised model PLANSUM (Amplayo et al., 2021b). We classify PLANSUM as a weakly-supervised summarizer since it uses additional information other than review text.
+
+A detailed introduction to the baselines is in Appendix B.
+
+Implementation Details We used RoBERTa-base (Liu et al., 2019) for the sentiment analysis, a BART-base (Lewis et al., 2020) encoder for the contrastive information valuation, and BART-base
+
+<table><tr><td rowspan="2">Summary</td><td colspan="3">Pros</td><td colspan="3">Cons</td><td colspan="3">Verdict</td></tr><tr><td>Info↑</td><td>Coh↑</td><td>N-R↑</td><td>Info↑</td><td>Coh↑</td><td>N-R↑</td><td>Info↑</td><td>Coh↑</td><td>N-R↑</td></tr><tr><td>GOLD</td><td>19.7</td><td>5.7</td><td>6.4</td><td>0</td><td>0</td><td>4</td><td>6.9</td><td>5.2</td><td>10.2</td></tr><tr><td>SELSUM</td><td>-20.2</td><td>2.9</td><td>11.8</td><td>-6.3</td><td>-0.6</td><td>-0.9</td><td>-7.2</td><td>6.5</td><td>0.6</td></tr><tr><td>BRIO</td><td>-7.5</td><td>-25.6</td><td>-23.5</td><td>-5.1</td><td>-7.3</td><td>-10.7</td><td>-8.4</td><td>-17.2</td><td>-14.7</td></tr><tr><td>SUBSUMM</td><td>8</td><td>17</td><td>5.3</td><td>11.4</td><td>7.9</td><td>7.6</td><td>8.6</td><td>5.5</td><td>3.9</td></tr></table>
+
+Table 3: Human evaluation results on AmaSum. Info, Coh, and N-R are abbreviations of Informativeness, Coherence, and Non-Redundancy.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">AmaSum</td><td rowspan="2">RT</td></tr><tr><td>Pros</td><td>Cons</td><td>Verdict</td></tr><tr><td>SUBSUMM</td><td>0.66</td><td>0.67</td><td>0.89</td><td>0.70</td></tr><tr><td>CHATGPT</td><td>0.34</td><td>0.33</td><td>0.11</td><td>0.30</td></tr></table>
+
+Table 4: Win rates of SUBSUMM and CHATGPT in the pair-wise comparisons.
+
+![](images/6db86effd5f195c18c642003105b89de7f94faeb4021410d42bc7ad45a5333c1.jpg)  
+Figure 2: A t-SNE (van der Maaten and Hinton, 2008) plot of the embeddings of most reviews from a product in AmaSum, where the dots of different colors represent the reviews selected for different perspectives. A few outliers are omitted.
+
+as the backbone of our summarizer and its variants. In all of the review sampling strategies, we selected  $K = 10$  reviews for every subset, which is explained in Appendix D. All the following experiments were conducted on 2 Geforce RTX 3090 GPUs. For the hyperparameters and more details, please refer to Appendix C.
+
+# 4.2 Results
+
+Automatic Evaluation We used ROUGE-1/2/L as the evaluation metrics and reported the F-1 scores. For AmaSum, we evaluated pros, cons, and verdict separately. As shown in Table 1 and Table 2, SUBSUMM significantly outperforms other methods on both datasets. Specifically, there are
+
+two observations:
+
+(1) SUBSUMM excels in generating summary with obvious emotion tendency, i.e., pros and cons. We notice that the three targets in AmaSum are equally treated by all the baselines, indicating the lack of exploration into the difference between the perspectives. SUBSUMM samples only positive/negative reviews for pros/cons summary. As depicted in Fig. 2, the reviews sampled for pros and cons are distributed in different zones of the semantic space, with the reviews for verdict evenly scattered between them. It not only reduces inconsistencies, but also adds valuable information to the input, since positive reviews always point out more advantages of the product, and vice versa.
+
+(2) The supervised methods score generally higher than the LLM-related methods, and SUBSUMM has an edge over the congener supervised systems. Although LLMs possess strong versatility in text generation, fine-tuning standard PLMs on annotated data seems non-trivial for opinion summarization. Compared with the other supervised methods, SUBSUMM reuses the reference summaries through contrastive learning in both information valuation and the training stage II, thus comprehensively utilizing the annotations. Especially, LONGFORMER's sparse attention mechanism plays an implicit selection role, while the review sampling strategies of SUBSUMM consider the sentiment tendency and information value, which is more sophisticated and task-specific.
+
+Human Evaluation As a supplement to automatic evaluation, we conducted a user study using Best-Worst Scaling (BWS; Louviere et al., 2015) detailed in Appendix E. The four evaluated summaries were GOLD (reference) summary and summaries generated by SELSUM, BRIO, and SUBSUMM. The three criteria were Informativeness, Coherence, and Non-Redundancy.
+
+Results in Table 3 demonstrate the considerable
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">Pros</td><td colspan="3">Cons</td><td colspan="3">Verdict</td></tr><tr><td>R-1</td><td>R-2</td><td>R-L</td><td>R-1</td><td>R-2</td><td>R-L</td><td>R-1</td><td>R-2</td><td>R-L</td></tr><tr><td colspan="10">Sampling Strategy</td></tr><tr><td>RAND</td><td>21.83</td><td>4.28</td><td>19.91</td><td>14.06</td><td>2.55</td><td>12.35</td><td>22.67</td><td>4.66</td><td>17.81</td></tr><tr><td>SENTI-RAND</td><td>22.33</td><td>4.41</td><td>20.45</td><td>14.92</td><td>2.75</td><td>13.13</td><td>23.06</td><td>4.79</td><td>17.98</td></tr><tr><td>SENTIINFO</td><td>22.82</td><td>4.60</td><td>20.80</td><td>14.42</td><td>2.48</td><td>12.59</td><td>23.51</td><td>5.03</td><td>18.16</td></tr><tr><td>SENTI-RANDINFO</td><td>23.01</td><td>4.76</td><td>21.02</td><td>15.46</td><td>2.92</td><td>13.60</td><td>23.51</td><td>5.04</td><td>18.32</td></tr><tr><td colspan="10">Training Scheme</td></tr><tr><td>SUBSUMM</td><td>26.25</td><td>4.96</td><td>24.18</td><td>16.72</td><td>3.00</td><td>14.80</td><td>25.36</td><td>5.04</td><td>19.58</td></tr><tr><td>w/o Stage I</td><td>25.48</td><td>4.58</td><td>23.50</td><td>16.65</td><td>2.94</td><td>14.60</td><td>24.93</td><td>4.78</td><td>19.44</td></tr><tr><td>w/o Stage II</td><td>23.01</td><td>4.76</td><td>21.02</td><td>15.46</td><td>2.92</td><td>13.60</td><td>23.51</td><td>5.04</td><td>18.32</td></tr><tr><td>RAND in Stage I</td><td>26.03</td><td>4.93</td><td>23.96</td><td>16.66</td><td>2.97</td><td>14.78</td><td>24.71</td><td>4.93</td><td>19.43</td></tr><tr><td>RAND in Stage II</td><td>25.95</td><td>4.95</td><td>24.02</td><td>16.50</td><td>2.77</td><td>14.54</td><td>24.60</td><td>4.96</td><td>19.45</td></tr></table>
+
+Table 5: Results of analysis experiments on AmaSum. R-1/2/L are the ROUGE-1/2/L F-1 scores, and the highest scores in both blocks are shown in bold.
+
+practical value of our model. Regarding Informativeness, summaries from SUBSUMM display comparable, even more information than GOLD summaries. As for Coherence, SUBSUMM leaves the users the best reading experience with correct grammar and straightforward expression. In terms of Non-Redundancy, SUBSUMM does not present the most succinct summaries, but considering the first two criteria, the redundancy is still acceptable.
+
+We further compared our model with the LLM via 50 head-to-head tests between SUBSUMM and CHATGPT. The test cases were randomly sampled from the two datasets (15 samples from each partition in AmaSum's test set and 5 samples from RT's test set), and the annotators were asked to make pair-wise comparisons without the reference summaries. The results are shown in Table 4. It seems that the users prefer the summaries generated by SUBSUMM. An obvious issue of CHATGPT is that it cannot control the output length within a few calls when the input is overlong. Consequently, most of the summaries generated are either excessively long or abruptly truncated with the maximum length argument fixed. In addition, though CHATGPT is qualified to produce fluent text, it suffers from more severe hallucination than our model, which may compromise its ROUGE scores. In Table 10 are some supporting cases.
+
+# 4.3 Analysis
+
+With the purpose of deeper insights into our proposal, we carried out some in-depth analysis exper
+
+iments on AmaSum, using the same metrics as in automatic evaluation. We also reported the results on RT in Appendix F.
+
+Comparison between Sampling Strategies As aforementioned, the quality of the three strategies in the review sampling strategy set would elevate sequentially. To confirm, we compare summarizers from the training stage I with different sampling strategies in the upper block of Table 5. RAND, SENTI-RAND, and SENTI-INFO apply Random Sampling, Sentiment-Random Sampling, and Sentiment-Information Ranking in training and inference respectively; SENTI-RAND-INFO is trained with Sentiment-Random Sampling but infers on review subsets produced by Sentiment-Information Ranking.
+
+By comparing RAND with SENTI-RAND, it can be seen that with the aid of sentiment analysis, the review subsets sampled appear more useful for the summaries with emotion tendencies. There is no clear improvement from SENTI-RAND to SENTI-INFO, so we add SENTI-RAND-INFO to ascertain the reason. SENTI-RAND-INFO and SENTI-RAND only differ in the test input, while the former wins with a clear margin, suggesting Sentiment-Information Ranking produces better review subsets. SENTI-RAND-INFO shares the same test input with SENTI-INFO but results in higher ROUGE scores, possibly because the stochastic factor prevents the potential over-fitting problem. It also drops a hint that employing diverse review subsets might promote the model performance.
+
+Insight into Two-Stage Training Scheme We investigate the gains from the two-stage training scheme through an ablation study. The variants in the bottom block of Table 5 share the same test input with SUBSUMM.
+
+Our experiments evidence that both stage I and stage II are significant to model performance, while the latter plays a greater role. We suppose that the two stages are complementary to each other: the standard MLE training in stage I functions as task-specific initialization, and the multi-task learning in stage II passes on more knowledge to the model, mitigating the exposure bias problem.
+
+Moreover, we explore how the two-stage training scheme contributes to the summary quality by replacing the sub-optimal and optimal strategies in the two training stages with Random Sampling. It leads to observable performance decreases, yet they are slighter than those when stage I or II is directly removed. It can be inferred that besides the complementary training objectives and additional training steps, the sensible selection of the review subsets is also conducive to model training.
+
+# 5 Conclusion
+
+In this paper, we put forward a supervised summarization framework for large-scale and multiperspective opinion summarization, SUBSUMM. SUBSUMM supports a two-stage training scheme based on a set of review sampling strategies of multiple quality levels. Our model surpasses the state-of-the-art models and LLM-related systems on AmaSum and RT, manifesting its superiority in dealing with plentiful reviews and displaying various points of view. The analysis experiments verify that both components of SUBSUMM help the summarizer achieve better results.
+
+In the future, we are planning to (1) explore more review sampling strategies to fully learn the aspect information and (2) combine the proposed framework with LLMs and generalize it to other large-scale multi-input tasks.
+
+# Limitations
+
+There are also some limitations in SUBSUMM. In Table 1, the verdict partition, the ROUGE-2 F1-score of our model does not outweigh that of SELSUM; the ROUGE-L F1-score of our model is slightly lower than that of CHATGPT.
+
+Firstly, since ROUGE-2 reflects 2-gram recall, we suspect that this is due to the absence of explicit
+
+designs for aspect learning in SUBSUMM, which causes the model to miss more 2-gram aspect terms than SELSUM (We noticed that SELSUM emphasizes aspect learning). Secondly, ROUGE-L is computed based on the longest common subsequence, which has something to do with the fluency of the generation. We find that there are some errors, like repetitions and incomplete first words in the summaries from SUBSUMM. Compared to the LLMs with extensive parameters, our proposal still has room for improvement in language modeling.
+
+# Ethics Statement
+
+We used only publicly available datasets, artifacts, and figures. Nevertheless, we realize that the proposed framework may produce fabricated and potentially harmful contents, for the PLMs used are pre-trained on heterogeneous web corpora. Therefore, we recommend that the users cautiously apply the proposal and its by-products.
+
+# Acknowledgements
+
+The work is partially supported by the National Nature Science Foundation of China (No. 61976160, 61906137, 61976158, 62076184, 62076182) and Shanghai Science and Technology Plan Project (No. 21DZ1204800) and Technology Research Plan Project of Ministry of Public and Security (Grant No. 2020JSYJD01).
+
+# References
+
+Ojas Ahuja, Jiacheng Xu, Akshay Gupta, Kevin Horecka, and Greg Durrett. 2022. ASPECTNEWS: Aspect-oriented summarization of news documents. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6494-6506, Dublin, Ireland. Association for Computational Linguistics.  
+Reinald Kim Amplayo, Stefanos Angelidis, and Mirella Lapata. 2021a. Aspect-controllable opinion summarization. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6578-6593, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Reinald Kim Amplayo, Stefanos Angelidis, and Mirella Lapata. 2021b. Unsupervised opinion summarization with content planning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 12489-12497.  
+Reinald Kim Amplayo and Mirella Lapata. 2020. Unsupervised opinion summarization with noising and
+
+denoising. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1934-1945, Online. Association for Computational Linguistics.  
+Iz Beltagy, Matthew E. Peters, and Arman Cohan. 2020. Longformer: The long-document transformer.  
+Adithya Bhaskar, Alexander R. Fabbri, and Greg Durrett. 2023. Prompted opinion summarization with gpt-3.5.  
+Arthur Bražinskas, Mirella Lapata, and Ivan Titov. 2020a. Few-shot learning for opinion summarization. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4119-4135, Online. Association for Computational Linguistics.  
+Arthur Bražinskas, Mirella Lapata, and Ivan Titov. 2020b. Unsupervised opinion summarization as copycat-review generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5151-5169, Online. Association for Computational Linguistics.  
+Arthur Bražinskas, Mirella Lapata, and Ivan Titov. 2021. Learning opinion summarizers by selecting informative reviews. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9424-9442, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Arthur Brazinskas, Ramesh Nallapati, Mohit Bansal, and Markus Dreyer. 2022. Efficient few-shot finetuning for opinion summarization. In *Findings of the Association for Computational Linguistics: NAACL* 2022, pages 1509-1523, Seattle, United States. Association for Computational Linguistics.  
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc.  
+Shuyang Cao and Lu Wang. 2021. CLIFF: Contrastive learning for improving faithfulness and factuality in abstractive summarization. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6633-6649, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.
+
+Eric Chu and Peter Liu. 2019. MeanSum: A neural model for unsupervised multi-document abstractive summarization. In Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 1223-1232. PMLR.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Sergey Edunov, Myle Ott, Michael Auli, David Grangier, and Marc'Aurelio Ranzato. 2018. Classical structured prediction losses for sequence to sequence learning. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 355-364, New Orleans, Louisiana. Association for Computational Linguistics.  
+Günes Erkan and Dragomir R. Radev. 2004. Lexrank: Graph-based lexical centrality as salience in text summarization. J. Artif. Int. Res., 22(1):457-479.  
+Kavita Ganesan, ChengXiang Zhai, and Jiawei Han. 2010. Opinosis: A graph based approach to abstractive summarization of highly redundant opinions. In Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 340-348, Beijing, China. Coling 2010 Organizing Committee.  
+Suyu Ge, Jiaxin Huang, Yu Meng, and Jiawei Han. 2023. Finesum: Target-oriented, fine-grained opinion summarization. In Proceedings of the Sixteenth ACM International Conference on Web Search and Data Mining, WSDM '23, page 1093-1101, New York, NY, USA. Association for Computing Machinery.  
+Jinbae Im, Moonki Kim, Hoyeop Lee, Hyunsouk Cho, and Sehee Chung. 2021. Self-supervised multimodal opinion summarization. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 388-403, Online. Association for Computational Linguistics.  
+Hayate Iso, Xiaolan Wang, Yoshihiko Suhara, Stefanos Angelidis, and Wang-Chiew Tan. 2021. Convex Aggregation for Opinion Summarization. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 3885-3903, Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Masaru Isonuma, Junichiro Mori, Danushka Bollegala, and Ichiro Sakata. 2021. Unsupervised abstractive
+
+opinion summarization by generating sentences with tree-structured topic guidance. Transactions of the Association for Computational Linguistics, 9:945-961.  
+Wenjun Ke, Jinhua Gao, Huawei Shen, and Xueqi Cheng. 2022. Consistsum: Unsupervised opinion summarization with the consistency of aspect, sentiment and semantic. In Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining, WSDM '22, page 467-475, New York, NY, USA. Association for Computing Machinery.  
+Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd International Conference for Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings.  
+Svetlana Kiritchenko and Saif Mohammad. 2017. Bestworst scaling more reliable than rating scales: A case study on sentiment intensity annotation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 465-470, Vancouver, Canada. Association for Computational Linguistics.  
+Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics.  
+Chunyuan Li, Xiang Gao, Yuan Li, Baolin Peng, Xiujun Li, Yizhe Zhang, and Jianfeng Gao. 2020. Optimus: Organizing sentences via pre-trained modeling of a latent space. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4678-4699, Online. Association for Computational Linguistics.  
+Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain. Association for Computational Linguistics.  
+Yang Liu and Mirella Lapata. 2019. Text summarization with pretrained encoders. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3730-3740, Hong Kong, China. Association for Computational Linguistics.  
+Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized BERT pretraining approach. CoRR, abs/1907.11692.
+
+Yixin Liu and Pengfei Liu. 2021. SimCLS: A simple framework for contrastive learning of abstractive summarization. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 1065-1072, Online. Association for Computational Linguistics.  
+Yixin Liu, Pengfei Liu, Dragomir Radev, and Graham Neubig. 2022. BRIO: Bringing order to abstractive summarization. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2890-2903, Dublin, Ireland. Association for Computational Linguistics.  
+Ilya Loshchilov and Frank Hutter. 2017. Sgdr: Stochastic gradient descent with restarts. In 5th International Conference for Learning Representations, ICLR 2017, Palais des Congrès Neptune, Toulon, France, 2017, Conference Track Proceedings.  
+Jordan Louviere, Terry Flynn, and A. A. J. Marley. 2015. Best-Worst Scaling: Theory, Methods and Applications. Cambridge University Press.  
+Ziming Mao, Chen Henry Wu, Ansong Ni, Yusen Zhang, Rui Zhang, Tao Yu, Budhaditya Deb, Chenguang Zhu, Ahmed Awadallah, and Dragomir Radev. 2022. DYLE: Dynamic latent extraction for abstractive long-input summarization. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1687-1698, Dublin, Ireland. Association for Computational Linguistics.  
+Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013. Distributed representations of words and phrases and their compositionality. In Advances in Neural Information Processing Systems, volume 26. Curran Associates, Inc.  
+OpenAI. 2023. Gpt-4 technical report.  
+Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. 2019. *fairoseq: A fast, extensible toolkit for sequence modeling*. In *Proceedings of NAACL-HLT* 2019: Demonstrations.  
+Nadav Oved and Ran Levy. 2021. PASS: Perturb-and-select summarizer for product reviews. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 351-365, Online. Association for Computational Linguistics.  
+Bo Pang, Erik Nijkamp, Wojciech Kryscinski, Silvio Savarese, Yingbo Zhou, and Caiming Xiong. 2023. Long document summarization with top-down and bottom-up inference. In *Findings of the Association for Computational Linguistics: EACL* 2023, pages 1267-1284, Dubrovnik, Croatia. Association for Computational Linguistics.
+
+Romain Paulus, Caiming Xiong, and Richard Socher. 2017. A deep reinforced model for abstractive summarization. CoRR, abs/1705.04304.  
+Ofir Press and Lior Wolf. 2017. Using the output embedding to improve language models. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 157-163, Valencia, Spain. Association for Computational Linguistics.  
+Alec Radford, Rafal Jozefowicz, and Ilya Sutskever. 2017. Learning to generate reviews and discovering sentiment. CoRR, abs/1704.01444.  
+Gaetano Rossiello, Pierpaolo Basile, and Giovanni Semeraro. 2017. Centroid-based text summarization through compositionality of word embeddings. In Proceedings of the MultiLing 2017 Workshop on Summarization and Summary Evaluation Across Source Types and Genres, pages 12-21, Valencia, Spain. Association for Computational Linguistics.  
+Ori Shapira and Ran Levy. 2020. Massive multi-document summarization of product reviews with weak supervision.  
+Shichao Sun and Wenjie Li. 2021. Alleviating exposure bias via contrastive learning for abstractive text summarization. CoRR, abs/2108.11846.  
+Laurens van der Maaten and Geoffrey E. Hinton. 2008. Visualizing high-dimensional data using t-sne. Journal of Machine Learning Research, 9:2579-2605.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc.  
+Ke Wang and Xiaojun Wan. 2021. TransSum: Translating aspect and sentiment embeddings for self-supervised opinion summarization. In *Findings of the Association for Computational Linguistics: ACLIJCNLP* 2021, pages 729-742, Online. Association for Computational Linguistics.  
+Lu Wang and Wang Ling. 2016. Neural network-based abstract generation for opinions and arguments. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 47-57, San Diego, California. Association for Computational Linguistics.  
+Ronald J. Williams. 1992. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine Learning, 8(3):229-256.  
+Shusheng Xu, Xingxing Zhang, Yi Wu, and Furu Wei. 2021. Sequence level contrastive learning for text summarization. CoRR, abs/2109.03481.
+
+Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, and Amr Ahmed. 2020. Big bird: Transformers for longer sequences. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS'20, Red Hook, NY, USA. Curran Associates Inc.  
+Wen Zhang, Yang Feng, Fandong Meng, Di You, and Qun Liu. 2019. Bridging the gap between training and inference for neural machine translation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4334-4343, Florence, Italy. Association for Computational Linguistics.  
+Ming Zhong, Pengfei Liu, Yiran Chen, Danqing Wang, Xipeng Qiu, and Xuanjing Huang. 2020. Extractive summarization as text matching. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6197-6208, Online. Association for Computational Linguistics.
+
+<table><tr><td>AmaSum</td><td>Train</td><td>Dev</td><td>Test</td></tr><tr><td>Entity</td><td>26,660</td><td>3,302</td><td>3,362</td></tr><tr><td>Reviews per Ent.</td><td>77.78</td><td>77.76</td><td>76.80</td></tr><tr><td>Tokens per Rev.</td><td>53.98</td><td>54.13</td><td>53.97</td></tr><tr><td>Tokens per Sum.</td><td></td><td></td><td></td></tr><tr><td>- Pros</td><td>39.13</td><td>39.21</td><td>38.81</td></tr><tr><td>- Cons</td><td>18.34</td><td>18.31</td><td>17.81</td></tr><tr><td>- Verdict</td><td>21.72</td><td>21.97</td><td>21.59</td></tr><tr><td>Rotten Tomatoes</td><td>Train</td><td>Dev</td><td>Test</td></tr><tr><td>Entity</td><td>2,453</td><td>536</td><td>737</td></tr><tr><td>Reviews per Ent.</td><td>74.39</td><td>74.85</td><td>73.26</td></tr><tr><td>Tokens per Rev.</td><td>26.52</td><td>26.32</td><td>26.41</td></tr><tr><td>Tokens per Sum.</td><td>23.80</td><td>23.59</td><td>23.82</td></tr></table>
+
+# A Dataset Statistics
+
+The statistics of the datasets after preprocessing are shown in Table 6. The average numbers are taken except for "Entity".
+
+# B Baselines
+
+On AmaSum dataset, the baselines are unsupervised extractive models (a) LEXRANK (Erkan and Radev, 2004), a PageRank-like algorithm that extracts sentences based on graph centrality, (b) EXTSUM (Bražinskas et al., 2021), which uses the same ROUGE greedy heuristic as in Liu and Lapata (2019); unsupervised abstractive models (c) MEANSUM (Chu and Liu, 2019), which generates opinion summary by reconstructing the mean of review embeddings, (d) COPYCAT (Bražinskas et al., 2020b), a VAE summarizer with hierarchical continuous latent representations to model products and individual reviews; supervised abstractive models (e) SELSUM, a model jointly learns to select informative subsets of reviews and summarizing the opinions, (f) LONGFORMER, a long-range model with an attention mechanism that scales linearly with sequence length, (g) BRIO, the state-of-the-art model of general abstractive summarization; and LLM-related solutions (h) GPT-3.5-turbo, (i) QG (Bhaskar et al., 2023), a pipeline where reviews are summarized by QFSumm (Ahuja et al., 2022) and GPT-3 (Brown et al., 2020), specifically the text-curie-001 model successively.
+
+On RT dataset are some other baselines: unsupervised extractive models (j) W2VCENT (Rossiello et al., 2017), (k) SNCENT (Amplayo and Lapata,
+
+Table 6: Statistics of the preprocessed datasets. The average numbers are taken except for "Entity".  
+
+<table><tr><td rowspan="2">Hyperparameter</td><td colspan="3">AmaSum</td><td rowspan="2">RT</td></tr><tr><td>Pros</td><td>Cons</td><td>Verdict</td></tr><tr><td>bsz I</td><td>16</td><td>16</td><td>16</td><td>8</td></tr><tr><td>bsz II</td><td>16</td><td>16</td><td>16</td><td>16</td></tr><tr><td>warmup I</td><td>5,000</td><td>5,000</td><td>5,000</td><td>500</td></tr><tr><td>warmup II</td><td>3,000</td><td>3,000</td><td>3,000</td><td>300</td></tr><tr><td>γ (Eq. 13)</td><td>0.1</td><td>1.0</td><td>0.1</td><td>0.1</td></tr><tr><td>lenpen</td><td>0.5</td><td>0.5</td><td>1.0</td><td>1.0</td></tr><tr><td>minlen</td><td>35</td><td>25</td><td>25</td><td>30</td></tr></table>
+
+Table 7: Hyperparameter setting.  $bsz$  I, II denote the batch sizes in the two training stages. minlen stands for the minimum generation length in the beam search algorithm.
+
+2020), and (l) BERTCENT (Amplayo et al., 2021b), which take encodings from word2vec (Mikolov et al., 2013), LSTM-based model (Radford et al., 2017), and BERT (Devlin et al., 2019) as the input representations; unsupervised abstractive models (m) OPINOSIS (Ganesan et al., 2010), a graph-based model that leverages token-level redundancy to summarize text, (n) DENOISESUM (Amplayo and Lapata, 2020), which re-formulates the summarization task as a denoising task; and a weakly supervised model (o) PLANSUM (Amplayo et al., 2021b), which constructs the synthetic dataset with a Dirichlet distribution parametrized by a content planner.
+
+# C Implementation Details
+
+The codes we used for fine-tuning the PLMs in sentiment analysis, contrastive information valuation, and the training stage I were implemented with Fairseq (Ott et al., 2019) library. For sentiment analysis, we set the learning rate to 1e-05 and updated the model parameters with the Adam optimizer. For contrastive information valuation, the margin  $\lambda$  in Eq. 7 was 1e-02. We set the learning rate to 3e-05 and adopted the Adam optimizer with the cosine learning rate scheduler (Loshchilov and Hutter, 2017). The minimum and maximum learning rates were 1e-08 and 3e-05.
+
+During the training stage I, the input reviews were independently encoded, and the concatenated hidden states of the reviews were attended by the decoder to predict the summary. Following Press and Wolf (2017), the token embeddings were shared across the encoder and decoder for regularization. We used the learning rate of 3e-05 and
+
+![](images/fadb88d509b045ad41a1239f62d341cb47622e940b22e2875f65017aaca74d35.jpg)  
+Figure 3: Prompts for LLMs in our experiments. The four prompts were used for pros, cons, and verdict in AmaSum and Rotten Tomatoes in order.
+
+![](images/68522be1a2b3b72bfd19fb97e46ad237b975d27b2951e8e423af6c99b5298385.jpg)
+
+the Adam optimizer (Kingma and Ba, 2015) with 5,000 warmup steps for model optimization. The decoding strategy was beam search with a beam size of 5 and trigram blocking (Paulus et al., 2017).
+
+In the training stage II, the reviews in the subset were joined with the separator  $<s>$  before being fed to the model, and  $M = 16$  candidate summaries were collected for every training sample. We used the Adam optimizer with the same learning rate scheduler as BRIO and changed the maximum learning rate to 1e-03. We used the margin  $\lambda$  of 1e-03 in Eq. 11 for all experiments. For summary generation, we used beam search of size 5. Particularly, We distinguished between the length penalty hyperparameter  $\alpha$  in Eq. 12 and lenpen in the beam search algorithm: the former was fixed at 2.0, and the latter differed across the targets. Other detailed hyperparameters are listed in Table 7.
+
+For the baselines, we adapted BRIO to the task of large-scale opinion summarization. Concretely, while preparing the input, we always sampled  $R_{1:N}$  out of  $R_{1:N}$  and joined the  $K$  reviews with the separator  $< s >$ . The input of BRIO and CHATGPT was the optimal review subset as in Sec. 3.2 due to the maximum input length of 4096; LONGFORMER (LED-base-16384) and QG received the original review set as input. The prompts for the LLMs are presented in Fig. 3.
+
+# D Hyperparameter Selection
+
+The hyperparameter  $K$  has a significant impact on the performance of SUBSUM. In this paper, we followed the setting of the baseline model SELSUM and inherited  $K = 10$  on AmaSum dataset for comparability. Before working on RT dataset, we had conducted experiments with varying  $K$  and random review subsets to explore the best value.
+
+<table><tr><td>Value</td><td>R-1</td><td>R-2</td><td>R-L</td></tr><tr><td>K=6</td><td>21.36</td><td>4.30</td><td>15.87</td></tr><tr><td>K=8</td><td>22.16</td><td>4.89</td><td>16.62</td></tr><tr><td>K=10</td><td>23.20</td><td>5.56</td><td>17.28</td></tr><tr><td>K=12</td><td>22.54</td><td>5.40</td><td>16.58</td></tr></table>
+
+Table 8: Results of the experiments with varying  $K$  and random review subsets on RT dataset. The highest scores are shown in bold.
+
+From the results in Table 8, it can be inferred that a too-small value of  $K$  can cause information deficiency, and a too-large one may introduce the sparsity problem even after the review selection, so we didn't change the value of  $K$ .
+
+# E Human Evaluation
+
+BWS is known to produce more reliable results than raking scales (Kiritchenko and Mohammad, 2017) and is widely used in opinion summarization studies. We randomly selected 30 samples from the pros, cons, and verdict partition of AmaSum's test set severally and recruited 6 volunteers. The volunteers were asked to choose one best and one worst summary from four summaries for three criteria and report the confidence of their choices. For each volunteer's response, the best model received  $+1$ , the worst model received  $-1$ , and the rest of the models received 0 scores. Taking the confidence as weight, the scores of 6 volunteers were weighted and summed to get the final scores.
+
+About the criteria, Informativeness tells if the summary presents opinions about specific aspects of the entity in the round, Coherence measures how easy the summary is to read and reflects if the summary follows a natural ordering of facts,
+
+<table><tr><td>Method</td><td>R-1</td><td>R-2</td><td>R-L</td></tr><tr><td colspan="4">Sampling Strategy</td></tr><tr><td>RAND</td><td>22.73</td><td>5.27</td><td>16.80</td></tr><tr><td>SENTI-RAND</td><td>22.76</td><td>5.36</td><td>16.81</td></tr><tr><td>SENTIINFO</td><td>23.44</td><td>5.84</td><td>17.25</td></tr><tr><td>SENTI-RANDINFO</td><td>23.99</td><td>6.09</td><td>17.54</td></tr><tr><td colspan="4">Training Scheme</td></tr><tr><td>SUBSUMM</td><td>24.96</td><td>6.66</td><td>19.08</td></tr><tr><td>w/o Stage I</td><td>23.72</td><td>5.16</td><td>18.05</td></tr><tr><td>w/o Stage II</td><td>23.99</td><td>6.09</td><td>17.54</td></tr><tr><td>RAND in Stage I</td><td>25.40</td><td>6.74</td><td>19.13</td></tr><tr><td>RAND in Stage II</td><td>25.46</td><td>6.77</td><td>18.90</td></tr></table>
+
+Table 9: Results of analysis experiments on RT. R-1/2/L are the ROUGE-1/2/L F-1 scores, and the highest scores in both blocks are shown in bold.
+
+and Non-Redundancy measures the repetitions and unnecessary contents in the summary.
+
+# F Experiment Results
+
+The results of the analysis experiments on RT are reported in Table 9. We list a set of example summaries from SUBSUMM and other baselines on the AmaSum dataset in Table 10.
+
+<table><tr><td>Method</td><td>Pros</td><td>Cons</td><td>Verdict</td></tr><tr><td>GOLD</td><td>A quality pumice stone crafted from 100% pure volcanic lava. Has a good shape, making it easy to hold and apply pressure to hard skin. Use the handy rope loop to hang the stone out to dry after use.</td><td>Some users find the coarse-ness rough and takes time to get used to.</td><td>Comfortable to hold and use, this all natural vol- canic pumice stone works hard on calluses and tough skin.</td></tr><tr><td>SELSUM</td><td>Made of natural materi- als that are easy to clean. Comes with a rope at- tached to hang in the shower for easy use.</td><td>Some users find it difficult to remove dead skin while it&#x27;s in use. May not re- move all dead skin on the feet, but it&#x27;s a good choice for those with dry skin.</td><td>A good choice if you&#x27;re looking for a natural pumice stone that&#x27;s easy to hold and use.</td></tr><tr><td>BRIO</td><td>A before and after, it was a bit embarrassing, and my feet were a few deep cracks for a few scrubs my feet are gone and a couple of scrubs</td><td>Users report that the pumice stone is not as smooth as they would like. Some reports of the stone breaking.</td><td>Pumice stone is the real deal and I appreciate that. using it for the first time is a great experience.</td></tr><tr><td>CHATGPT</td><td>Pumice Valley&#x27;s natural pumice stone is effective in removing dead skin, comes with a handy rope for easy use, and has a fine grain that doesn&#x27;t rip</td><td>The stone is not sharp or rough enough to re- move dry skin, it may be chipped or in poor shape, it can</td><td>The pumice stone is effec- tive in removing dead skin and smoothing heels, but may not work well on ex- isting calluses.</td></tr><tr><td>SUBSUMM</td><td>Natural lava stone that&#x27;s easy to use and easy to clean. Comes with a rope for hanging in the shower. Comes in a variety of sizes. Made from lava.</td><td>May be too rough for some users. Some users find it difficult to remove the calluses from the stone. May not remove all cal- luses.</td><td>You&#x27;re looking for a pumice stone that&#x27;s easy to use and clean, this is the one to buy.</td></tr></table>
+
+Table 10: Example summaries from SUBSUMM and other baselines on AmaSum. Contents that coincide with the reference summaries or appear erroneous for opinion summarization are highlighted.

2023/Large-Scale and Multi-Perspective Opinion Summarization with Diverse Review Subsets/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:724c378041206dcd5eb5610b497f2a76112915fc1f65b37ff9c6e1108b3152de
+size 896004

2023/Late Fusion of Transformers for Sentiment Analysis of Code-Switched Data/61718999-dcc2-4475-b2ab-d1733664f853_content_list.json

@@ -0,0 +1,1018 @@
+[
+    {
+        "type": "text",
+        "text": "Late Fusion of Transformers for Sentiment Analysis of Code-Switched Data",
+        "text_level": 1,
+        "bbox": [
+            115,
+            89,
+            880,
+            110
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Gagan Sharma and R Chinmay and Raksha Sharma",
+        "bbox": [
+            265,
+            136,
+            737,
+            153
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "gagan_s@cs.iitr.ac.in and r_c@cs.iitr.ac.in and raksha.sharma@cs.iitr.ac.in",
+        "bbox": [
+            186,
+            154,
+            815,
+            170
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Indian Institute of Technology Roorke",
+        "bbox": [
+            347,
+            171,
+            653,
+            187
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Roorkee, Uttarakhand, India",
+        "bbox": [
+            384,
+            187,
+            616,
+            202
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Code-switching is a common phenomenon in multilingual communities and is often used on social media. However, sentiment analysis of code-switched data is a challenging yet less explored area of research. This paper aims to develop a sentiment analysis system for code-switched data. In this paper, we present a novel approach combining two transformers using logits of their output and feeding them to a neural network for classification. We show the efficacy of our approach using two benchmark datasets, viz., English-Hindi (En-Hi), and English-Spanish (En-Es) availed by Microsoft GLUECoS. Our approach results in an  $F1$  score of 73.66% for En-Es and 61.24% for En-Hi, significantly higher than the best model reported for the GLUECoS benchmark dataset.",
+        "bbox": [
+            141,
+            282,
+            460,
+            537
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            552,
+            258,
+            569
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Code-Switching refers to using two or more languages in a single conversation. This linguistic feature is prevalent in multilingual communities all over the world. Multilingual communities share vast amounts of unstructured code-switched data on social media in text and speech. Sentiment analysis of data shared on social media has proved valuable in market predictions and analysis (Darapaneni et al., 2022), mental health assessments (Calvo et al., 2017), and intelligent agents (Rambocas and Pacheco, 2018). There have been enormous studies conducted on multilingual NLP, but, despite such prevalence of code-switching in real life, very few studies have been conducted in the field.",
+        "bbox": [
+            112,
+            580,
+            489,
+            804
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Wang et al., (2018) proposed Generalized Language Evaluation Benchmark (GLUE) to evaluate embedding models on various language understanding tasks. Inspired by this, Microsoft Research published GLUECoS (Khanuja et al., 2020), an evaluation benchmark on various tasks focused on code-switched text. Sentiment Analysis is one of",
+        "bbox": [
+            112,
+            806,
+            489,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "table",
+        "img_path": "images/fc70ee5a718e0587b2137898e2838ffecabb0d1edf904d1a8e7bcdd93987fbed.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Hindi</td><td>Romanized</td></tr><tr><td>न्दय (new)</td><td>naya nga</td></tr><tr><td>रैशटम (end)</td><td>khatam khtm khatm</td></tr><tr><td>बस (sufficient)</td><td>bas bs bus</td></tr></table>",
+        "bbox": [
+            579,
+            256,
+            805,
+            312
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Figure 1: Hindi words with multiple possible Romanization",
+        "bbox": [
+            507,
+            329,
+            884,
+            357
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "the seven tasks present in the GLUECoS evaluation benchmark, which is aimed at classifying sentiment in a code-switched sentence as positive, negative, or neutral. The task consists of English-Spanish and English-Hindi code-switched language pairs. The English-Spanish task is useful for studying the performance of models in a code-switched environment of closely related languages, and the English-Hindi task is for largely different languages.",
+        "bbox": [
+            507,
+            385,
+            884,
+            530
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The code-switched data analysis comes with its challenges. Traditional techniques trained for one language (monolingual NLP models) quickly break down when input is mixed in from another language. The performance of these models goes down as the amount of mixed-language present increases (Aguilar et al., 2019). On the other hand, there is a scarcity of high-quality code-switched datasets (Sitaram et al., 2020). Conversations on social media usually use slang, unknown references, mentions, improper grammar, spelling errors, etc. Models trained on correct grammar and spelling will not perform well on such datasets immediately (Dhungra et al., 2016).",
+        "bbox": [
+            507,
+            531,
+            882,
+            755
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In addition to this, there is also the case of multiple Romanization. A code switch statement is usually found in the Roman script because of standard English keyboards' vast popularity and availability. Multiple Romanization leads to words in foreign languages written in English alphabets with varying spellings, as shown in English-Hindi examples in Figure 1.",
+        "bbox": [
+            507,
+            757,
+            884,
+            885
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The cause of such different spellings is usually the difference in pronunciations which gets",
+        "bbox": [
+            507,
+            887,
+            882,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "6485",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 6485-6490",
+        "bbox": [
+            216,
+            945,
+            779,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            277,
+            957,
+            719,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "reflected in the absence or addition of vowels in the Romanized foreign words.",
+        "bbox": [
+            112,
+            84,
+            485,
+            116
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our contribution to the sentiment analysis of code-switched data research includes the performance improvement from the combination (late fusion) of different pre-trained models without using ensemble techniques. Since pre-trained models are trained on different datasets and have different architectures, they produce varying results on the same dataset. However, a unison model, which includes multiple pre-trained models, each fine-tuned on whole training data, improves performance compared to ensemble or other voting techniques. We show that our unisom model has improved the weighted-F1 score by  $10\\%$  for the English-Spanish about  $3.73\\%$  for the English-Hindi sentiment analysis task.",
+        "bbox": [
+            112,
+            118,
+            489,
+            357
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The rest of the paper is organized as follows. In Section 2, we briefly describe the details regarding the datasets used. Section 3 describes some recent related work. Section 4 describes our approach. Section 5 provides the experimental setup. Section 6 lists the results obtained, and Section 7 concludes the paper.",
+        "bbox": [
+            112,
+            361,
+            489,
+            474
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Dataset Description",
+        "text_level": 1,
+        "bbox": [
+            112,
+            489,
+            319,
+            507
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "GLUECoS is a reasonably significant evaluation benchmark for the six code-switched NLP tasks. A high performance sentiment analysis system for this dataset opens the scope for improvement in the performance of other code-switched NLP tasks. In this paper, we present results on the GLUECoS benchmark dataset. The dataset is generated by running a script that scrapes tweets of some preserved ids annotated beforehand (Khanuja et al., 2020). The English-Spanish dataset consists of 1523 tweets for training, 184 for validation, and 232 for testing. The English-Hindi dataset consists of 10080 tweets for training, 1260 for validation, and 1261 for testing. We have referred to the best results and method (mBERT) reported on the GLUECoS website<sup>1</sup> for the task. The objective of the paper is to improve upon the state-of-the-art method, that is, mBERT adapted for code-switched sentiment analysis.",
+        "bbox": [
+            112,
+            518,
+            489,
+            824
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Every data item is labeled to be expressing a positive, negative, and neutral sentiment. The detailed distribution of the annotations are listed in Table 1 and 2. Below is the definition of assigned labels.",
+        "bbox": [
+            112,
+            826,
+            489,
+            891
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "- Positive: The statement infers joy, encouragement, likes, sarcasm, or excitement. For example, I am amazed by the work done by the VFX team.",
+        "bbox": [
+            531,
+            84,
+            882,
+            148
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "- Negative: The statement infers sadness, discouragement, dislikes, sadism, or indifference. For example, The food service was horrible at our hotel.",
+        "bbox": [
+            531,
+            162,
+            884,
+            227
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "- Neutral: The statement is mostly factual and does not demonstrate much of an opinion. For example, The Taj Mahal is one of the World's Seven Wonders.",
+        "bbox": [
+            531,
+            242,
+            882,
+            306
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "table",
+        "img_path": "images/4289e453e1fe2e6b03e32add5552f7bcad603598031a5f696f96ee38346b76d1.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Sentiment</td><td>En-Es</td><td>En-Hi</td></tr><tr><td>Positive</td><td>433</td><td>3202</td></tr><tr><td>Negative</td><td>379</td><td>2319</td></tr><tr><td>Neutral</td><td>711</td><td>4559</td></tr><tr><td>Total</td><td>1523</td><td>10080</td></tr></table>",
+        "bbox": [
+            576,
+            332,
+            818,
+            419
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "table",
+        "img_path": "images/2ce762cf25d281bf60fa3abba214e730c0a639e20f43a41d1cc4e5ac815162ee.jpg",
+        "table_caption": [
+            "Table 1: Training Set Distribution"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Sentiment</td><td>En-Es</td><td>En-Hi</td></tr><tr><td>Positive</td><td>55</td><td>399</td></tr><tr><td>Negative</td><td>46</td><td>283</td></tr><tr><td>Neutral</td><td>83</td><td>578</td></tr><tr><td>Total</td><td>184</td><td>1260</td></tr></table>",
+        "bbox": [
+            576,
+            472,
+            816,
+            558
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Validation Set Distribution",
+        "bbox": [
+            573,
+            567,
+            816,
+            581
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 Literature Survey",
+        "text_level": 1,
+        "bbox": [
+            507,
+            617,
+            697,
+            634
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Several research and industrial works have been done on sentiment analysis on textual data, especially in the previous decade. Recently, the research focused on code-switched data has been growing in popularity since a large share of posts available in social media or general conversations includes two or more languages (Winata et al., 2022). Before 2019, most sentiment analysis research revolved around techniques such as lexical sentiment analysis (like Bag-of-Words technique) (Kasthuriarachchy et al., 2014), semi-supervised or unsupervised algorithms (Maas et al., 2011), convolution models (Jianqiang et al., 2018), recurrent models (Arras et al., 2017), etc. These models failed to capture the context of the sentences and needed to be more accurate compared to transformers (Ran et al., 2018). Transformer-based models",
+        "bbox": [
+            505,
+            645,
+            884,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>https://microsoft.github.io/GLUECoS/",
+        "bbox": [
+            136,
+            903,
+            379,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "6486",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "replaced these classical models since the availability of labeled data was the main bottleneck for training neural networks (Devlin et al., 2019).",
+        "bbox": [
+            112,
+            84,
+            489,
+            131
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Semi-supervised algorithms such as Label Propagation (LP) algorithms (Wang et al., 2015) were used to predict sentiments in Code-switched data, using both statistical machine translation and sentiment information for words. In 2016, Wang et al., (2016) proposed a Bilingual Attention Network for code-switched emotion Prediction. They used Bi-LSTM to capture the context of words and both monolingual and bilingual attention vectors to predict the overall emotion of the input sentence. However, transformers (Vaswani et al., 2017) and pre-trained language models produced much better results than traditional techniques (Luo and Wang, 2019), with significantly less training data available. Fine-tuning the models with small amounts of available labeled data produced much better results than earlier methods. Recent research includes the generation of synthetic code-switched data (for better fine-tuning), intermediate task training (Prasad et al., 2021), continued pre-training (Li et al., 2022) (masked Language modeling using code-switched texts), late fusion (Mundra et al., 2021) (using multiple pre-trained transformer-based models and combining their outputs using some logic to improve overall performance), and Data Augmentation (Mundra et al., 2021) - both Random Augmentation (RA) and Balanced Augmentation (BA), and custom attention models (Li et al., 2021). Other techniques are used to improve sentiment analysis results in code-switched texts specific to language pairs or dataset specifications. In this paper, we propose a novel approach to enhance the late fusion technique using logits from models instead of the output.",
+        "bbox": [
+            115,
+            133,
+            490,
+            682
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4 Methodology",
+        "text_level": 1,
+        "bbox": [
+            112,
+            697,
+            265,
+            714
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Due to the complementary behavior of different sentiment prediction models, we use the late fusion (Colneric and Demar, 2020) technique to combine results from different models as a unison model. The unison model, which includes multiple pretrained models, each fine-tuned on whole training data, improves performance compared to ensemble or other voting techniques. We combine multiple models using their logit prediction for each sentiment. The description of such a model is given in Figure 2. The logit values for each sentiment class generated from two models are fed to a neural",
+        "bbox": [
+            112,
+            726,
+            490,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "network with three nodes in the output layer, one for each sentiment class. The neural network post-model training handles assigning weights to the results generated by each model. It thus gives more weight to the models as and when one performs better than the other and thus produce a new model which has the better aspects of each of the contributing model. The output neuron at the output layer with the highest value determines the sentiment class for the given input.",
+        "bbox": [
+            507,
+            84,
+            884,
+            244
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Logit is described as the output of a neuron without applying the activation function. Logit for a class is defined using the probability of the class being the predicted output. The mathematical definition of the logit is provided below.",
+        "bbox": [
+            507,
+            244,
+            884,
+            325
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL o g i t (p) = l o g (\\frac {p}{(1 - p)})\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            598,
+            332,
+            794,
+            362
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "It is clear from the mathematical description that a larger logit value means a larger probability for the said sentiment class.",
+        "bbox": [
+            507,
+            367,
+            882,
+            414
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "5 Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            507,
+            426,
+            717,
+            443
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "This section provides the experimental details to replicate the work.",
+        "bbox": [
+            507,
+            451,
+            880,
+            482
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "5.1 Data Pre-processing",
+        "text_level": 1,
+        "bbox": [
+            507,
+            493,
+            715,
+            508
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The dataset contains words that are written in different scripts. For the English-Hindi task, we need to transliterate words written in the Devanagari script to the Roman script to make it suitable for further processing. This is done using libindic's indic-trans (Bhat et al., 2015) library.",
+        "bbox": [
+            507,
+            514,
+            882,
+            609
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "5.2 Hyper-parameters of the models",
+        "text_level": 1,
+        "bbox": [
+            507,
+            620,
+            811,
+            636
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We have implemented BERT (Devlin et al., 2019), mBERT, XLM-R (Conneau et al., 2019) models in our work. For English-Spanish Task, the hyperparameters are mentioned in Table 3. We used AdamW Optimizer (Loshchilov and Hutter, 2019) for all models. Other values are default values from the Hugging-Face library (refer BERT, mBERT, XLM-R).",
+        "bbox": [
+            507,
+            640,
+            884,
+            769
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/0f09e2f7203a332d48d1abedb99623e06f1ec6f648937f233d9b24c31fc6379e.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>gain</td><td>epsilon</td><td>batch</td><td>epoch</td></tr><tr><td>BERT</td><td>5e-5</td><td>1e-8</td><td>16</td><td>5</td></tr><tr><td>mBERT</td><td>4e-5</td><td>1e-8</td><td>16</td><td>5</td></tr><tr><td>XLM-R</td><td>3e-5</td><td>1e-8</td><td>16</td><td>5</td></tr></table>",
+        "bbox": [
+            524,
+            778,
+            867,
+            847
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Hyper-parameters for En-Es models",
+        "bbox": [
+            541,
+            857,
+            848,
+            872
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 4 mentions the parameters for the English-Hindi task. For this task, we have used Hi-BERT",
+        "bbox": [
+            507,
+            887,
+            884,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "6487",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/891e428313d7336a8c97832328877a9ba022bdf93fc05cda872fae529a758292.jpg",
+        "image_caption": [
+            "Figure 2: Late Fusion Neural Network"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            260,
+            83,
+            736,
+            186
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "which is an ELECTRA model (Clark et al., 2020) trained on Hindi Wikipedia data and Hindi CommonCrawl dataset. For tokenization, we used Bert-Tokenizer. Other values are the default from the Hugging-Face library (refer Hi-BERT).",
+        "bbox": [
+            112,
+            235,
+            490,
+            317
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/012a1af8e4ef28eeecd8591f9d252b84b40e977826a748acc57c9e0f2b30291c.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>gain</td><td>epsilon</td><td>batch</td><td>epoch</td></tr><tr><td>BERT</td><td>5e-5</td><td>1e-8</td><td>32</td><td>3</td></tr><tr><td>mBERT</td><td>4e-5</td><td>1e-8</td><td>32</td><td>3</td></tr><tr><td>XLM-R</td><td>3e-5</td><td>1e-8</td><td>16</td><td>5</td></tr><tr><td>Hi-BERT</td><td>5e-5</td><td>1e-8</td><td>32</td><td>3</td></tr></table>",
+        "bbox": [
+            124,
+            326,
+            478,
+            412
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "5.3 Evaluation Metric",
+        "text_level": 1,
+        "bbox": [
+            112,
+            464,
+            305,
+            479
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We used the weighted  $F1$  score as a metric of performance measure. The evaluation is done through a pull request at the GLUECoS GitHub repository. GitHub Actions run an evaluation script and generates the score.",
+        "bbox": [
+            112,
+            485,
+            490,
+            565
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "5.4 Late Fusion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            577,
+            253,
+            590
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The two models generate logits for each sentiment class. These values are fed to a neural network with six input neurons (Figure 2). Further, the model has one hidden layer with six hidden neurons and an output layer with three output neurons (one for each sentiment category). The neural network is created using the MLPClassifier of SKlearn. Other values are the default from the SKlearn implementation.",
+        "bbox": [
+            112,
+            599,
+            489,
+            728
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "6 Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            740,
+            213,
+            753
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "BERT has proven to be state-of-the-art for many NLP tasks, including sentiment analysis. For sentiment analysis of code-switched data, a multilingual version of BERT, that is, mBERT (Khanuja et al., 2020), is reported as state-of-the-art for the GLUE-CoS benchmark dataset. In this paper, we have tried to present a more accurate system by making an ensemble of multiple BERT-based models",
+        "bbox": [
+            112,
+            766,
+            490,
+            894
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/fcc1fcc50471d48e96d0b8be09ed85bf929300810ebd6ec4cd8c7e889de5df34.jpg",
+        "table_caption": [
+            "Table 4: Hyper-parameters for En-Hi models"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>F1</td></tr><tr><td>mBERT (Khanuja et al., 2020)</td><td>63.02%</td></tr><tr><td>BERT</td><td>65.05%</td></tr><tr><td>XLM-R</td><td>69.42%</td></tr><tr><td>mBERT - BERT (Late Fusion)</td><td>69.40%</td></tr><tr><td>XLM-R - BERT (Late Fusion)</td><td>70.9%</td></tr><tr><td>XLM-R - mBERT (Late Fusion)</td><td>73.66%</td></tr></table>",
+        "bbox": [
+            517,
+            231,
+            873,
+            350
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 5: F1 Scores Obtained for English-Spanish Sentiment Analysis.",
+        "bbox": [
+            507,
+            357,
+            882,
+            388
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "suitable for multiple languages text (viz., mBERT, XLM-R, Hi-BERT). Tables 5 and 6 present the results of our approach and various other methods for English-Spanish and English-Hindi code-switched dataset for sentiment analysis. Khanuja et al., (2020) presented mBERT as these datasets' best-performing model. We further produced results with BERT, XLM-R for both the datasets and Hi-BERT for the English-Hindi dataset. The results obtained from the late fusion of two models are better than those obtained from individual models. In our late fusion approach, the neural network post-model training handles assigning weights to the results generated by each model. It thus gives more weight to the models as and when one performs better than the other and thus produce a new model which has the better aspects of each of the contributing model. The late fusion technique reduces the predictions' dispersion, making the end model more robust. We obtained the best F1 score of  $73.66\\%$  for English-Spanish with the late fusion of XLM-R and mBERT. While for English-Hindi, mBERT and BERT produced the best F1 score of  $61.24\\%$ .",
+        "bbox": [
+            507,
+            434,
+            884,
+            819
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "3The majority voting-based ensemble method results were worse than the independent use of BERT/mBERT/XLM-R models. Hence, we focused on showing improvement over the individual model.",
+        "bbox": [
+            507,
+            869,
+            882,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "2refer MLPClassifier.",
+        "bbox": [
+            134,
+            903,
+            272,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "6488",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/295cefc7287d3abc92208c3173f447e2438ec8400cbde0a46782c017bdcbc7d0.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>F1</td></tr><tr><td>XLM-R</td><td>28.5%</td></tr><tr><td>Hi-BERT</td><td>55.90%</td></tr><tr><td>mBERT (Khanuja et al., 2020)</td><td>57.51%</td></tr><tr><td>BERT</td><td>59.05%</td></tr><tr><td>BERT - Hi-BERT (Late Fusion)</td><td>59.73%</td></tr><tr><td>mBERT - Hi-BERT (Late Fusion)</td><td>59.97%</td></tr><tr><td>mBERT - BERT (Late Fusion)</td><td>61.24%</td></tr></table>",
+        "bbox": [
+            124,
+            80,
+            477,
+            214
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 6: F1 Scores Obtained for English-Hindi Sentiment Analysis.",
+        "bbox": [
+            112,
+            223,
+            489,
+            253
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "7 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            114,
+            278,
+            245,
+            293
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "This paper presents a novel approach that uses late fusion for sentiment analysis of code-switched data. The results show that mBERT-BERT and XLMR-mBERT fusion models perform the best for English-Hindi and English-Spanish Sentiment Analysis tasks, respectively. The tools and models used in our work are largely non-language specific and thus could be helpful in Sentiment Analysis of code-switched data of different languages.",
+        "bbox": [
+            112,
+            305,
+            487,
+            450
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            462,
+            218,
+            476
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "This paper presents results for code-switched data in English-Hindi and English-Spanish. The languages involved in the work are widely spoken languages across the world over social media. However, we have not observed the performance of our fusion-based approach for code-switched data where one of the languages is low resource language such as English and Dravidian (Malayalam, Tamil, Kannada). It requires further investigation and experimentation.",
+        "bbox": [
+            112,
+            488,
+            489,
+            649
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Ethics Statement",
+        "text_level": 1,
+        "bbox": [
+            114,
+            662,
+            265,
+            678
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The work reported in the paper complies with the ACL Ethics Policy. The output of our work does not have any kind of negative impact on society; it is for the good of society.",
+        "bbox": [
+            112,
+            688,
+            487,
+            753
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            114,
+            765,
+            285,
+            781
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We are grateful to Dr. Rudra Murthy from IBM Research Bangalore for his guidance throughout the execution of the project.",
+        "bbox": [
+            112,
+            791,
+            487,
+            840
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            868,
+            213,
+            883
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "ref_text",
+        "text": "Gustavo Aguilar, Fahad AlGhamdi, Victor Soto, Mona Diab, Julia Hirschberg, and Thamar Solorio. 2019.",
+        "bbox": [
+            114,
+            891,
+            487,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Named entity recognition on code-switched data: Overview of the calcs 2018 shared task.",
+            "Leila Arras, Grégoire Montavon, Klaus-Robert Müller, and Wojciech Samek. 2017. Explaining recurrent neural network predictions in sentiment analysis.",
+            "Irshad Ahmad Bhat, Vandan Mujadia, Aniruddha Tammewar, Riyadh Ahmad Bhat, and Manish Shrivastava. 2015. Iiit-h system submission for fire2014 shared task on transliterated search. In Proceedings of the Forum for Information Retrieval Evaluation, FIRE '14, pages 48-53, New York, NY, USA. ACM.",
+            "Rafael Alejandro Calvo, David N. Milne, M. Sazzad Hussain, and Helen Christensen. 2017. Natural language processing in mental health applications using non-clinical texts†. *Natural Language Engineering*, 23:649 - 685.",
+            "Kevin Clark, Minh-Thang Luong, Quoc V. Le, and Christopher D. Manning. 2020. Electra: Pre-training text encoders as discriminators rather than generators.",
+            "Niko Colneric and Janez Demar. 2020. Emotion recognition on twitter: Comparative study and training a unison model. IEEE Transactions on Affective Computing, 11:433-446.",
+            "Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Unsupervised cross-lingual representation learning at scale. In Annual Meeting of the Association for Computational Linguistics.",
+            "Narayana Darapaneni, Anwesh Reddy Paduri, Himank Sharma, Milind Manjrekar, Nutan Hindlekar, Pranali Bhagat, Usha Aiyer, and Yogesh Agarwal. 2022. Stock price prediction using sentiment analysis and deep learning for indian markets.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding.",
+            "Bhuwan Dhingra, Zhong Zhou, Dylan Fitzpatrick, Michael Muehl, and William W. Cohen. 2016. Tweet2vec: Character-based distributed representations for social media.",
+            "Zhao Jianqiang, Gui Xiaolin, and Zhang Xuejun. 2018. Deep convolution neural networks for twitter sentiment analysis. IEEE Access, 6:23253-23260.",
+            "Buddhika Kasthuriarachchy, Kasun De Zoysa, and Lalith Premaratne. 2014. Enhanced bag-of-words model for phrase-level sentiment analysis.",
+            "Simran Khanuja, Sandipan Dandapat, Anirudh Srinivasan, Sunayana Sitaram, and Monjit Choudhury. 2020. Gluecos: An evaluation benchmark for codeswitched nlp."
+        ],
+        "bbox": [
+            509,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "6489",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Bin Li, Yixuan Weng, Qiya Song, Bin Sun, and Shutao Li. 2022. Continuing pre-trained model with multiple training strategies for emotional classification. In Proceedings of the 12th Workshop on Computational Approaches to Subjectivity, Sentiment & Social Media Analysis, pages 233-238, Dublin, Ireland. Association for Computational Linguistics.",
+            "Ruifan Li, Hao Chen, Fangxiang Feng, Zhanyu Ma, Xiaojie Wang, and Eduard Hovy. 2021. Dual graph convolutional networks for aspect-based sentiment analysis. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 6319-6329, Online. Association for Computational Linguistics.",
+            "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization.",
+            "Li Luo and Yue Wang. 2019. Emotionx-hsu: Adopting pre-trained bert for emotion classification. ArXiv, abs/1907.09669.",
+            "Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. 2011. Learning word vectors for sentiment analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 142-150, Portland, Oregon, USA. Association for Computational Linguistics.",
+            "Jay Mundra, Rohan Gupta, and Sagnik Mukherjee. 2021. WASSA@IITK at WASSA 2021: Multi-task learning and transformer finetuning for emotion classification and empathy prediction. In Proceedings of the Eleventh Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis, pages 112-116, Online. Association for Computational Linguistics.",
+            "Archiki Prasad, Mohammad Ali Rehan, Shreya Pathak, and Preethi Jyothi. 2021. The effectiveness of intermediate-task training for code-switched natural language understanding. In Proceedings of the 1st Workshop on Multilingual Representation Learning, pages 176-190, Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Meena Rambocas and Barney Pacheco. 2018. Online sentiment analysis in marketing research: a review. Journal of Research in Interactive Marketing, 12.",
+            "Li Ran, Lianzhe Zheng, Lin Hailun, Wang Weiping, and Meng Dan. 2018. Text emotion analysis: A survey. Journal of Computer Research and Development, 55:30.",
+            "Sunayana Sitaram, Khyathi Raghavi Chandu, Sai Krishna Rallabandi, and Alan W Black. 2020. A survey of code-switched speech and language processing."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need.",
+            "Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353-355, Brussels, Belgium. Association for Computational Linguistics.",
+            "Zhongqing Wang, Sophia Lee, Shoushan Li, and Guodong Zhou. 2015. Emotion detection in code-switching texts via bilingual and sentimental information. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 763-768, Beijing, China. Association for Computational Linguistics.",
+            "Zhongqing Wang, Yue Zhang, Sophia Lee, Shoushan Li, and Guodong Zhou. 2016. A bilingual attention network for code-switched emotion prediction. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1624-1634, Osaka, Japan. The COLING 2016 Organizing Committee.",
+            "Winata, Aji, Yong, and Solorio. 2022. The decades progress on code-switching research in nlp: A systematic survey on trends and challenges."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            532
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "6490",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 5
+    }
+]